Deploy latest Objectverse Diary from fa09aac

.playwright-cli/console-2026-06-08T11-55-41-257Z.log

@@ -0,0 +1 @@
+[  101248ms] [ERROR] Failed to load resource: net::ERR_INCOMPLETE_CHUNKED_ENCODING @ http://127.0.0.1:7861/gradio_api/heartbeat/ybhxcw7jgrp:0

docs/07-development-plan.md

@@ -143,7 +143,7 @@ Verification:
 
 Goal: make persona, diary, and chat generation use a small local text model runtime.
 
-Status: optional runtime wiring complete; real GGUF smoke test pending.
+Status: optional runtime wiring complete; published LoRA v2 Q4_K_M GGUF passed local llama.cpp smoke. Hosted Space text runtime validation is still pending.
 
 Scope:
 
@@ -158,12 +158,12 @@ Scope:
 Exit criteria:
 
 - Text generation can run through llama.cpp or documented local fallback.
-- README documents runtime path. Final model size remains pending until GGUF selection.
+- README documents runtime path and published GGUF selection.
 - Trace records include runtime metadata.
 
 Verification:
 
-- Local runtime smoke test with a real GGUF.
+- Local runtime smoke test with `objectverse-diary-qwen15b-lora-v2-q4_k_m.gguf`.
 - JSON schema validation.
 - Compare at least three object generations for persona consistency.
 

docs/DATASET.md

@@ -4,7 +4,7 @@
 
 The project now has a deterministic SFT preview generator for local planning and schema validation.
 
-Current local artifact:
+Current preview artifact:
 
 ```bash
 .venv/bin/python -B scripts/generate_dataset.py
@@ -18,9 +18,9 @@ data/train/objectverse_sft_preview.jsonl
 
 This preview is mock-generated. It is not a final training dataset and should not be described as real model output.
 
-The stable submission baseline does not publish a final Hugging Face Dataset. The current JSONL file is evidence for schema and workflow readiness only.
+The preview JSONL file is evidence for schema and workflow readiness only.
 
-Additional local training-test artifact:
+Curated v1 training-test artifact:
 
 ```bash
 .venv/bin/python -B scripts/prepare_curated_dataset.py \
@@ -36,11 +36,22 @@ Published synthetic curated dataset:
 https://huggingface.co/datasets/qqyule/objectverse-diary-sft-curated
 ```
 
+Current curated v2 artifact:
+
+```bash
+.venv/bin/python -B scripts/prepare_curated_dataset.py \
+  --version v2 \
+  --count 200 \
+  --output data/train/objectverse_sft_curated_v2.jsonl
+```
+
+The published dataset repo now includes `objectverse_sft_curated_v2.jsonl`: 200 synthetic curated rows covering 40 everyday objects and 5 personality modes, with exactly 40 rows per mode and no repeated object-mode pair. The v1 file remains preserved through repository history.
+
 ## Target Dataset
 
-Final target before fine-tuning:
+Target before stronger fine-tuning:
 
-- 200-500 generated object-persona-diary samples
+- 200-500 generated or curated object-persona-diary samples
 - at least 50 manually curated high-quality samples
 - no private user photos
 - no emails, tokens, serial numbers, or other sensitive identifiers
@@ -82,7 +93,7 @@ Full candidate pool later:
 .venv/bin/python -B scripts/generate_dataset.py --count 300 --output data/train/objectverse_sft_candidates.jsonl
 ```
 
-Manual curation should happen after generation. Do not publish the full candidate file until it has been reviewed.
+Manual curation should happen after generation. For a stronger LoRA run, curate 150-300 rows from a broader object/mode/scene pool and leave 10-15% for evaluation. Do not publish the full candidate file until it has been reviewed.
 
 Space VLM validation traces under `data/traces/space-vlm/` are failure evidence because they include `vision-fallback-to-mock`. Do not mix them into curated training data or describe them as successful real VLM outputs.
 
@@ -105,27 +116,77 @@ Validate the local JSONL shape without Modal auth or GPU usage:
   --run-name objectverse-diary-qwen15b-curated-test
 ```
 
-Intended training command after explicit confirmation:
+The first badge-evidence run used 20 steps on 50 synthetic curated rows. For a higher-quality v2 run, validate the larger curated file first:
 
 ```bash
-modal run scripts/finetune_lora.py \
-  --dataset data/train/objectverse_sft_curated.jsonl \
-  --run-name objectverse-diary-qwen15b-curated-test \
-  --max-steps 20
+.venv/bin/python -B scripts/finetune_lora.py \
+  --dry-run \
+  --dataset data/train/objectverse_sft_curated_v2.jsonl \
+  --run-name objectverse-diary-qwen15b-lora-v2 \
+  --max-steps 120 \
+  --learning-rate 1e-4 \
+  --max-seq-length 1536 \
+  --lora-r 32 \
+  --lora-alpha 64 \
+  --per-device-train-batch-size 2 \
+  --gradient-accumulation-steps 4 \
+  --eval-ratio 0.1 \
+  --eval-steps 20
+```
+
+Executed v2 training command:
+
+```bash
+modal run --timestamps -n objectverse-diary-qwen15b-lora-v2 scripts/finetune_lora.py \
+  --dataset data/train/objectverse_sft_curated_v2.jsonl \
+  --run-name objectverse-diary-qwen15b-lora-v2 \
+  --max-steps 120 \
+  --learning-rate 1e-4 \
+  --max-seq-length 1536 \
+  --lora-r 32 \
+  --lora-alpha 64 \
+  --per-device-train-batch-size 2 \
+  --gradient-accumulation-steps 4 \
+  --eval-ratio 0.1 \
+  --eval-steps 20
 ```
 
-Current Modal status: the curated test job completed successfully and produced the published LoRA adapter at `https://huggingface.co/qqyule/objectverse-diary-qwen15b-lora`.
+Current Modal status: the v2 job completed successfully and produced the published LoRA adapter at `https://huggingface.co/qqyule/objectverse-diary-qwen15b-lora`.
+
+Current v2 run summary:
+
+- run name: `objectverse-diary-qwen15b-lora-v2`
+- dataset: `data/train/objectverse_sft_curated_v2.jsonl`
+- dataset repo path: `objectverse_sft_curated_v2.jsonl`
+- records: 200 total, 180 train, 20 eval
+- base model: `Qwen/Qwen2.5-1.5B-Instruct`
+- max steps: 120
+- learning rate: `1e-4`
+- max sequence length: 1536
+- LoRA rank / alpha / dropout: 32 / 64 / 0.05
+- effective batch size: 8
+- assistant-output-only loss: enabled
+- train loss: 0.3240
+- eval loss: 0.0162
+- train runtime: 140.3364s
+- epoch: 5.2222
+- local adapter export: ignored `exports/objectverse-diary-qwen15b-lora-v2-adapter-dir/`
+- model repo: `https://huggingface.co/qqyule/objectverse-diary-qwen15b-lora`
+
+Additional v2 scaffold validation run: `objectverse-diary-qwen15b-lora-v2-curated50-retry1` completed on Modal with the existing 50-row curated dataset, using assistant-output-only loss, 45 train rows, 5 eval rows, `max_steps=120`, `learning_rate=1e-4`, `max_seq_length=1536`, LoRA `r=32`, `alpha=64`, and effective batch size 8. Metrics: `train_loss=0.2551`, `eval_loss=0.0093`, `train_runtime=146.5398s`, `epoch=20.0`. The adapter was downloaded to ignored local `exports/`; it has not been published to Hugging Face Hub.
 
 Default training scaffold settings:
 
 - base model: `Qwen/Qwen2.5-1.5B-Instruct`
 - LoRA adapter target: persona and diary JSON output
+- default loss: assistant-output-only labels, with prompt tokens masked
+- default eval split: 10% when the dataset has at least two rows
 - GPU: Modal `A10G`
 - output: Modal Volume artifacts, not committed files
 
 The current `objectverse_sft_preview.jsonl` file is mock-generated and should only be used to validate the training pipeline. It is not final Well-Tuned evidence. Do not store Modal credit codes, tokens, Hugging Face tokens, or private datasets in the repo.
 
-The published `objectverse_sft_curated.jsonl` dataset is synthetic curated training-test data. It is suitable for hackathon training evidence, but it should still be described honestly as a small synthetic set rather than real user trace data.
+The published `objectverse_sft_curated_v2.jsonl` dataset is synthetic curated training data. It is suitable for hackathon training evidence, but it should still be described honestly as deterministic synthetic curation rather than real user trace data.
 
 ## Curation Checklist
 
@@ -139,10 +200,20 @@ The published `objectverse_sft_curated.jsonl` dataset is synthetic curated train
 
 ## Publishing Notes
 
-Before publishing to Hugging Face Datasets:
+When publishing to Hugging Face Datasets:
 
 - create a dataset card
 - document that mock preview rows are synthetic
 - separate curated rows from raw candidates
 - include license and privacy notes
 - keep private images out of the repo
+
+Curated v2 was published with:
+
+```bash
+.venv/bin/python -B scripts/publish_hf_dataset.py \
+  --dataset-file data/train/objectverse_sft_curated_v2.jsonl \
+  --repo-id qqyule/objectverse-diary-sft-curated \
+  --path-in-repo objectverse_sft_curated_v2.jsonl \
+  --commit-message "Upload Objectverse Diary curated v2 dataset"
+```

docs/DEMO_VIDEO_SCRIPT.md

@@ -4,7 +4,7 @@
 
 Record a 90-second stable demo for Objectverse Diary using the mock-safe Hugging Face Space or local Gradio app.
 
-Do not claim that GGUF text generation or live LoRA runtime wiring are complete. Hosted MiniCPM-V validation is complete for the vision path, but the stable demo should still emphasize the mock-safe product loop, Gradio Off-Brand UI, public traces, published dataset/LoRA evidence, and no commercial AI APIs.
+Do not claim that live Space LoRA/GGUF runtime wiring is complete. Hosted MiniCPM-V validation is complete for the vision path and local GGUF text smoke is complete, but the stable demo should still emphasize the mock-safe product loop, Gradio Off-Brand UI, public traces, published dataset/model evidence, and no commercial AI APIs.
 
 ## Recording Setup
 
@@ -104,6 +104,6 @@ Screen:
 ## Notes For Submission
 
 - Mention MiniCPM-V as hosted-validated for object understanding, while the public demo defaults to mock for reliability.
-- Mention the published synthetic curated dataset and LoRA adapter only as training evidence, not live Space runtime.
+- Mention the published synthetic curated dataset, LoRA adapter, and Q4_K_M GGUF as model evidence, not live Space runtime.
 - Mention public traces and failure notes if the submission form asks for reproducibility.
 - Keep the final video under 2 minutes.

docs/DEVELOPMENT_STATUS.md

@@ -38,21 +38,22 @@ Last updated: 2026-06-08
   - social post draft
   - stable submission guide
 - Well-Tuned evidence:
-  - 50-row synthetic curated SFT dataset published at https://huggingface.co/datasets/qqyule/objectverse-diary-sft-curated
-  - Modal Qwen 1.5B LoRA test run completed with 20 steps
-  - LoRA adapter published at https://huggingface.co/qqyule/objectverse-diary-qwen15b-lora
-- GGUF smoke-test helper:
+  - 200-row synthetic curated v2 SFT dataset published at https://huggingface.co/datasets/qqyule/objectverse-diary-sft-curated
+  - Modal Qwen 1.5B LoRA v2 run completed with 120 steps, 180 train rows, and 20 eval rows
+  - LoRA v2 adapter published at https://huggingface.co/qqyule/objectverse-diary-qwen15b-lora
+- LoRA v2 GGUF runtime evidence:
   - `scripts/check_llama_cpp_smoke.py`
-  - recommended baseline model documented as `Qwen/Qwen2.5-1.5B-Instruct-GGUF` / `qwen2.5-1.5b-instruct-q4_k_m.gguf`
+  - adapter merged into `Qwen/Qwen2.5-1.5B-Instruct`
+  - pinned `llama.cpp` commit: `8f83d6c271d194bde2d410145a0ce73bc42e85cd`
+  - published Q4_K_M GGUF: https://huggingface.co/qqyule/objectverse-diary-qwen15b-lora/blob/main/objectverse-diary-qwen15b-lora-v2-q4_k_m.gguf
+  - local smoke passed with `llama-cpp text generation`, schema-valid persona/diary, non-empty chat, and no `text-fallback-to-mock`
   - trace runtime no longer records literal `TEXT_MODEL_PATH`
 - Local tests and initial acceptance currently pass.
 
 ## Not Completed
 
-- Real GGUF download/configuration outside Git and `TEXT_MODEL_PATH` smoke test. Model selection is now documented, but the file is not downloaded and optional `llama-cpp-python` is not installed by default.
-- Final text model parameter count documentation.
-- Real text model traces from non-mock runtime.
-- GGUF conversion and runtime wiring for the published LoRA adapter.
+- Hosted Space text runtime validation with the published GGUF. The public Space still uses mock-safe text until this passes.
+- Real text model traces from the hosted non-mock text runtime.
 - Published Field Notes URL, recorded demo video URL, social post URL, and final public submission.
 
 ## Current Safe Defaults
@@ -68,6 +69,15 @@ For a stable public baseline, keep the mock-safe Space as the demo path and only
 
 Next model gate:
 
+Download or mount the published GGUF on the target runtime, set:
+
+```bash
+OBJECTVERSE_TEXT_BACKEND=llama-cpp
+TEXT_MODEL_PATH=/absolute/path/to/objectverse-diary-qwen15b-lora-v2-q4_k_m.gguf
+```
+
+Then rerun the local or Space smoke path before claiming live text runtime.
+
 Optional rerun gate if Space variables, secrets, or dependencies change:
 
 ```bash

docs/FAILURES.md

@@ -12,10 +12,10 @@ MiniCPM-V 2.6 is wired as an optional vision backend. Hosted Space ZeroGPU valid
 
 The app includes a hidden `/vision_runtime_probe` API and `scripts/check_space_vlm.py` writes probe output into the Space VLM report before image validation. This probe identified the previous failure as a gated-model access issue rather than a GPU or dependency issue.
 
-The recommended baseline GGUF for local text smoke testing is selected, but not downloaded or run:
+The published LoRA v2 GGUF for local text smoke testing is available and has passed local llama.cpp smoke:
 
-- repo: `Qwen/Qwen2.5-1.5B-Instruct-GGUF`
-- file: `qwen2.5-1.5b-instruct-q4_k_m.gguf`
+- repo: `qqyule/objectverse-diary-qwen15b-lora`
+- file: `objectverse-diary-qwen15b-lora-v2-q4_k_m.gguf`
 - helper: `scripts/check_llama_cpp_smoke.py`
 
 Known non-blocking warning:
@@ -70,15 +70,27 @@ Known non-blocking warning:
 - Fallback used: mock object understanding plus mock text runtime if validation reaches generation.
 - Resolution: unresolved; keep the public Space mock-safe until this section reports a passing VLM validation.
 
-## 2026-06-08 - GGUF Smoke Helper Prepared, Actual Smoke Pending
+## 2026-06-08 - LoRA v2 GGUF Local Smoke Passed
 
 - Area: llama.cpp text runtime evidence.
-- Reproduction: Run `scripts/check_llama_cpp_smoke.py` with an external GGUF model path after optional dependency installation.
+- Reproduction: Run `scripts/check_llama_cpp_smoke.py` with `models/objectverse-diary-qwen15b-lora-v2-q4_k_m.gguf` after optional `llama-cpp-python` installation.
 - Expected: trace records `llama-cpp text generation`, persona/diary/chat run without `text-fallback-to-mock`.
-- Actual: not run; `.venv` does not include `llama-cpp-python` by default and the GGUF file is intentionally not committed.
-- Impact: Llama Champion evidence remains incomplete.
-- Fallback used: default mock text runtime remains the safe public demo path.
-- Resolution: pending explicit confirmation to install optional local dependency and download `qwen2.5-1.5b-instruct-q4_k_m.gguf` into ignored `models/`.
+- Actual: passed locally; trace included only `mock-vision-runtime` because text was real and vision remained mock for the smoke input.
+- Impact: local llama.cpp text runtime evidence is ready. Public Space text runtime is still not validated with this GGUF.
+- Fallback used: none for text.
+- Resolution: resolved locally by using the merged LoRA v2 Q4_K_M GGUF and conservative JSON extraction / decoding settings.
+- Evidence: `scripts/check_llama_cpp_smoke.py`, `docs/RUNTIME.md`, and the Hub file in `qqyule/objectverse-diary-qwen15b-lora`.
+
+## 2026-06-08 - Hugging Face Xet GGUF Upload Stalled
+
+- Area: Hugging Face model file upload.
+- Reproduction: Upload `models/objectverse-diary-qwen15b-lora-v2-q4_k_m.gguf` with the default Hub client path.
+- Expected: upload completes and commits `objectverse-diary-qwen15b-lora-v2-q4_k_m.gguf`.
+- Actual: the first upload stalled with Xet TLS EOF / `CLOSE_WAIT` after partial progress.
+- Impact: upload needed a retry; local GGUF file was unaffected.
+- Fallback used: stopped the stalled upload process and retried with `HF_HUB_DISABLE_XET=1`.
+- Resolution: resolved; ordinary Hub/LFS upload succeeded.
+- Evidence: Hub file `https://huggingface.co/qqyule/objectverse-diary-qwen15b-lora/blob/main/objectverse-diary-qwen15b-lora-v2-q4_k_m.gguf`.
 
 ## Anticipated Failure Areas
 

docs/FIELD_NOTES.md

@@ -137,10 +137,10 @@ For text runtime evidence, the project now includes a local smoke helper for an
 
 ```bash
 .venv/bin/python -B scripts/check_llama_cpp_smoke.py \
-  --model-path models/qwen2.5-1.5b-instruct-q4_k_m.gguf
+  --model-path models/objectverse-diary-qwen15b-lora-v2-q4_k_m.gguf
 ```
 
-The recommended baseline file is `qwen2.5-1.5b-instruct-q4_k_m.gguf` from `Qwen/Qwen2.5-1.5B-Instruct-GGUF`. It is intentionally not committed.
+The published local-smoke file is `objectverse-diary-qwen15b-lora-v2-q4_k_m.gguf` from `qqyule/objectverse-diary-qwen15b-lora`. It is intentionally not committed. Local smoke passed on June 8, 2026; Space text runtime still needs a separate validation before it should be described as live.
 
 ## 10. Privacy And Safety
 
@@ -150,12 +150,12 @@ Trace logging anonymizes text inputs before public export. The current public tr
 
 ## 11. What I Would Improve Next
 
-The next model-focused step is to smoke-test a real GGUF text model through llama.cpp.
+The next model-focused step is to validate the published GGUF in the hosted Space runtime, or keep it as local llama.cpp evidence while the public demo remains mock-safe.
 
 After that:
 
-- run the documented GGUF smoke test after explicit confirmation
-- decide whether the published LoRA should remain badge evidence only or be converted later
+- download or mount the published GGUF in the target runtime
+- set `OBJECTVERSE_TEXT_BACKEND=llama-cpp` and `TEXT_MODEL_PATH` for that runtime
 - generate real non-mock traces if hosted/local model validation passes
 - record a final demo video from the stable Space
 

docs/FINAL_VERIFICATION_REPORT.md

@@ -8,15 +8,16 @@
 
 ## Summary
 
-Objectverse Diary's stable mock-safe baseline remains locally verifiable. This update adds non-secret MiniCPM-V runtime diagnostics through a hidden Gradio API, probe-aware Space VLM reporting, a latest-failure-note updater, and a local llama.cpp GGUF smoke-test helper.
+Objectverse Diary's stable mock-safe baseline remains locally verifiable. This update adds non-secret MiniCPM-V runtime diagnostics through a hidden Gradio API, probe-aware Space VLM reporting, a latest-failure-note updater, and local llama.cpp GGUF smoke-test support. A later local run merged the LoRA v2 adapter, produced a Q4_K_M GGUF, uploaded it to the model repo, and passed local llama.cpp smoke.
 
-This report does not claim real GGUF text generation, live LoRA runtime wiring, Field Notes publication, demo video publication, social post publication, or final public submission URLs are complete.
+This report does not claim live Space LoRA/GGUF runtime wiring, Field Notes publication, demo video publication, social post publication, or final public submission URLs are complete.
 
 ## Implementation Additions
 
 - Hidden `/vision_runtime_probe` Gradio API returns sanitized backend, dependency, GPU, and MiniCPM-V load diagnostics.
 - `scripts/check_space_vlm.py` can include probe output in markdown/JSON reports and update the latest failure section in `docs/FAILURES.md`.
 - `scripts/check_llama_cpp_smoke.py` validates persona, diary, and chat through an externally configured GGUF without committing model files.
+- LoRA v2 GGUF tooling now covers merge, publish, and local smoke for `objectverse-diary-qwen15b-lora-v2-q4_k_m.gguf`.
 - Runtime status no longer records literal `TEXT_MODEL_PATH`; traces only record whether an external GGUF path is configured.
 - Submission docs now distinguish final-draft materials from published URLs.
 
@@ -88,11 +89,11 @@ No GGUF file, real token, private key, credential, or `.env` file was added by t
 - Demo video URL is still pending recording/publication.
 - Field Notes URL is still pending publication.
 - Social post URL is still pending publication.
-- Real GGUF download, optional `llama-cpp-python` installation, and smoke test remain pending explicit confirmation.
-- GGUF conversion and live runtime wiring for the published LoRA adapter remain future work.
+- Hosted Space text runtime validation with the published GGUF remains pending.
+- Live Space runtime wiring for the published LoRA/GGUF remains future work.
 
 ## Verdict
 
 PASS for the stable mock-safe local submission baseline plus local diagnostics/smoke-helper implementation.
 
-The project is ready for explicit-confirmation external steps: push `main`, sync the Space, rerun probe-aware Space VLM validation, run the local GGUF smoke test after optional dependency/model setup, record/publish the demo video, publish Field Notes/social post, and fill final submission URLs.
+The project is ready for explicit-confirmation external steps: push `main`, sync the Space, rerun probe-aware Space VLM validation if needed, validate the published GGUF in the Space runtime before claiming live text generation, record/publish the demo video, publish Field Notes/social post, and fill final submission URLs.

docs/MODEL_CARD.md

@@ -2,9 +2,9 @@
 
 ## Status
 
-Stable submission baseline plus one published text LoRA test adapter. The public Gradio Space still defaults to deterministic mock text; the adapter is training evidence and has not been converted to GGUF or wired into the live runtime.
+Stable submission baseline plus one published text LoRA v2 adapter and one published Q4_K_M GGUF. The public Gradio Space still defaults to deterministic mock text; the GGUF has passed local llama.cpp smoke, but it has not been switched into the live Space runtime.
 
-The app defaults to deterministic mock backends. MiniCPM-V 2.6 vision is wired as an optional runtime backend for GPU environments, with a hidden non-secret probe for hosted diagnostics. Text generation has optional llama.cpp wiring for an externally configured GGUF model via `TEXT_MODEL_PATH`. A Modal LoRA test run completed for the planned text model path and the adapter is published at `https://huggingface.co/qqyule/objectverse-diary-qwen15b-lora`.
+The app defaults to deterministic mock backends. MiniCPM-V 2.6 vision is wired as an optional runtime backend for GPU environments, with a hidden non-secret probe for hosted diagnostics. Text generation has optional llama.cpp wiring for an externally configured GGUF model via `TEXT_MODEL_PATH`. A Modal LoRA v2 run completed, the adapter is published at `https://huggingface.co/qqyule/objectverse-diary-qwen15b-lora`, and the merged Q4_K_M GGUF is published in the same repo.
 
 Hosted MiniCPM-V validation passed on June 8, 2026 after adding an `HF_TOKEN` Space secret with access to the gated `openbmb/MiniCPM-V-2_6` model. The validation used public mug, keyboard, and shoe images on ZeroGPU, while text generation intentionally remained mock. See `docs/SPACE_VLM_REPORT.md`.
 
@@ -19,8 +19,8 @@ Hosted MiniCPM-V validation passed on June 8, 2026 after adding an `HF_TOKEN` Sp
 | Component | Candidate | Notes |
 | --- | --- | --- |
 | Vision | `openbmb/MiniCPM-V-2_6` or mock fallback | Wired as optional backend; hosted ZeroGPU validation passed, then Space rolled back to mock-safe defaults. |
-| Text | deterministic mock text; published `Qwen/Qwen2.5-1.5B-Instruct` LoRA test adapter | Adapter published; not converted to GGUF or wired into Space runtime. |
-| Runtime | optional GGUF through llama.cpp / llama-cpp-python | Wired with mock fallback; smoke helper exists, real-model smoke test still pending. |
+| Text | deterministic mock text by default; published `Qwen/Qwen2.5-1.5B-Instruct` LoRA v2 Q4_K_M GGUF for local runtime | Adapter and GGUF published; Space text runtime remains mock-safe. |
+| Runtime | optional GGUF through llama.cpp / llama-cpp-python | Wired with mock fallback; local GGUF smoke passed on 2026-06-08. |
 | UI | Gradio Blocks | Required by the hackathon and project rules. |
 
 ## Parameter Budget
@@ -34,7 +34,7 @@ Record final numbers here before submission:
 | Vision | MiniCPM-V 2.6 optional path | ~8B | yes, when enabled |
 | Text base | Stable baseline mock text | 0 | no model parameters |
 | Optional text base | `Qwen/Qwen2.5-1.5B-Instruct` | ~1.5B | yes, when enabled |
-| Recommended GGUF smoke file | `Qwen/Qwen2.5-1.5B-Instruct-GGUF` / `qwen2.5-1.5b-instruct-q4_k_m.gguf` | ~1.5B base, quantized file | yes, if used for text runtime smoke |
+| Published LoRA v2 GGUF | `qqyule/objectverse-diary-qwen15b-lora` / `objectverse-diary-qwen15b-lora-v2-q4_k_m.gguf` | ~1.5B base, quantized file | yes, if enabled |
 | Published LoRA adapter | `qqyule/objectverse-diary-qwen15b-lora` | small adapter over base model | yes, when enabled |
 | Stable baseline total | Mock text + optional wired vision not active by default | 0 active model parameters by default | <= 32B |
 
@@ -63,7 +63,7 @@ Dataset planning lives in `docs/DATASET.md`.
 
 Current preview data is deterministic and mock-generated. It should only be used for schema validation, dry-run validation, and workflow planning until real candidate samples are generated and curated.
 
-The Modal training scaffold defaults to `Qwen/Qwen2.5-1.5B-Instruct` and saves adapter artifacts to a Modal Volume. `data/train/objectverse_sft_curated.jsonl` contains 50 synthetic curated rows for pipeline testing and is published at `https://huggingface.co/datasets/qqyule/objectverse-diary-sft-curated`.
+The Modal training scaffold defaults to `Qwen/Qwen2.5-1.5B-Instruct` and saves adapter artifacts to a Modal Volume. `data/train/objectverse_sft_curated_v2.jsonl` contains 200 synthetic curated rows covering 40 everyday objects and 5 personality modes. It is published at `https://huggingface.co/datasets/qqyule/objectverse-diary-sft-curated` as `objectverse_sft_curated_v2.jsonl`.
 
 Published adapter:
 
@@ -71,25 +71,32 @@ Published adapter:
 https://huggingface.co/qqyule/objectverse-diary-qwen15b-lora
 ```
 
-Training run summary:
+Current v2 training run summary:
 
 - Platform: Modal
-- Run name: `objectverse-diary-qwen15b-curated-test`
+- Run name: `objectverse-diary-qwen15b-lora-v2`
 - Base model: `Qwen/Qwen2.5-1.5B-Instruct`
-- Dataset: 50 synthetic curated rows
-- Steps: 20
-- Max sequence length: 1024
-- Learning rate: 0.0002
-- LoRA rank / alpha / dropout: 16 / 32 / 0.05
-- Train loss: 1.6697
-- GGUF conversion: not completed
+- Dataset: 200 synthetic curated v2 rows
+- Train / eval rows: 180 / 20
+- Steps: 120
+- Max sequence length: 1536
+- Learning rate: 0.0001
+- Effective batch size: 8
+- LoRA rank / alpha / dropout: 32 / 64 / 0.05
+- Assistant-output-only loss: enabled
+- Train loss: 0.3240
+- Eval loss: 0.0162
+- Epoch: 5.2222
+- GGUF conversion: completed with pinned `llama.cpp` commit `8f83d6c271d194bde2d410145a0ce73bc42e85cd`
+- Published GGUF: `objectverse-diary-qwen15b-lora-v2-q4_k_m.gguf`
 
 GGUF smoke status:
 
-- Recommended repo: `Qwen/Qwen2.5-1.5B-Instruct-GGUF`
-- Recommended file: `qwen2.5-1.5b-instruct-q4_k_m.gguf`
+- Repo: `qqyule/objectverse-diary-qwen15b-lora`
+- File: `objectverse-diary-qwen15b-lora-v2-q4_k_m.gguf`
 - Local helper: `scripts/check_llama_cpp_smoke.py`
-- Current state: file not downloaded, optional `llama-cpp-python` not installed by default, smoke test not run.
+- Local result: passed on 2026-06-08 with `llama-cpp text generation`, no `text-fallback-to-mock`, schema-valid persona and diary, and non-empty chat reply.
+- Space result: not run; do not claim live Space text runtime until a separate Space validation passes.
 
 ## Safety And Privacy
 

docs/SOCIAL_POST.md

@@ -5,8 +5,8 @@
 I built Objectverse Diary for Build Small Hackathon: a Gradio app where everyday objects wake up, get secret personas, write diaries, chat with you, and generate share cards.
 
 Stable demo: mock-safe, reproducible, no commercial AI APIs.
-MiniCPM-V hosted validation now passes for the vision path; llama.cpp is wired behind a local GGUF smoke helper.
-Synthetic curated dataset + Qwen 1.5B LoRA adapter are published as training evidence.
+MiniCPM-V hosted validation now passes for the vision path; local llama.cpp smoke passes with the published LoRA v2 Q4_K_M GGUF.
+Synthetic curated v2 dataset + Qwen 1.5B LoRA v2 adapter/GGUF are published as model evidence.
 
 Space: https://huggingface.co/spaces/build-small-hackathon/ObjectverseDiary
 
@@ -23,9 +23,9 @@ Objectverse Diary is my Build Small Hackathon project: a strange little object a
 - a shareable personality card
 - an anonymized trace record
 
-The stable submission baseline is mock-safe and reproducible, with no commercial AI APIs. MiniCPM-V vision is wired and hosted-validated on ZeroGPU, while llama.cpp text remains an optional local GGUF path.
+The stable submission baseline is mock-safe and reproducible, with no commercial AI APIs. MiniCPM-V vision is wired and hosted-validated on ZeroGPU, while llama.cpp text is validated locally through an optional GGUF path.
 
-I also published a small synthetic curated SFT dataset and a Qwen 1.5B LoRA test adapter for Well-Tuned evidence. The adapter is not wired into the public Space runtime yet; the live demo stays intentionally reliable.
+I also published a 200-row synthetic curated v2 SFT dataset, a Qwen 1.5B LoRA v2 adapter, and a Q4_K_M GGUF for model evidence. The GGUF is not wired into the public Space runtime yet; the live demo stays intentionally reliable.
 
 Space:
 https://huggingface.co/spaces/build-small-hackathon/ObjectverseDiary
@@ -38,4 +38,4 @@ https://huggingface.co/spaces/build-small-hackathon/ObjectverseDiary
 
 - Add GitHub URL after push is confirmed.
 - Add demo video URL after recording.
-- Do not claim GGUF smoke test or live LoRA runtime wiring are complete.
+- Do not claim live Space LoRA/GGUF runtime wiring is complete.

docs/SUBMISSION_GUIDE.md

@@ -24,6 +24,7 @@
 - Public mock traces: `data/traces/samples/`
 - Stable demo baseline: Gradio example buttons replay committed sample traces first, then fall back to the live generation pipeline if a cached trace is missing.
 - Optional llama.cpp runtime wiring: `src/models/llama_cpp_runner.py`
+- Published LoRA v2 Q4_K_M GGUF: https://huggingface.co/qqyule/objectverse-diary-qwen15b-lora/blob/main/objectverse-diary-qwen15b-lora-v2-q4_k_m.gguf
 
 ## Completed Locally
 
@@ -33,15 +34,16 @@
 - Optional llama.cpp text runtime wiring through `TEXT_MODEL_PATH`.
 - Hosted Space VLM validation script, report, JSON summary, and trace evidence export.
 - Hosted Space VLM probe support, latest failure-note update support, and passing MiniCPM-V ZeroGPU validation after adding an `HF_TOKEN` Space secret for gated model access.
-- Local GGUF smoke-test helper for `Qwen/Qwen2.5-1.5B-Instruct-GGUF` / `qwen2.5-1.5b-instruct-q4_k_m.gguf`; actual GGUF smoke remains pending.
-- Synthetic curated SFT dataset published to Hugging Face Datasets.
-- Modal Qwen 1.5B LoRA test run completed and adapter published to Hugging Face Models.
+- Local GGUF smoke-test helper passed with `models/objectverse-diary-qwen15b-lora-v2-q4_k_m.gguf`; trace text runtime was `llama-cpp text generation` and no `text-fallback-to-mock` was present.
+- Synthetic curated v2 SFT dataset published to Hugging Face Datasets: 200 rows, 40 objects, 5 personality modes.
+- Modal Qwen 1.5B LoRA v2 run completed and adapter published to Hugging Face Models.
+- LoRA v2 adapter merged into `Qwen/Qwen2.5-1.5B-Instruct`, converted with pinned `llama.cpp`, quantized to Q4_K_M, and uploaded to the same model repo.
 - Field Notes draft, demo video script, and social post draft for the stable submission package.
 
 ## Not Completed Yet
 
-- Real GGUF `TEXT_MODEL_PATH` smoke test and final text model parameter count. The recommended baseline GGUF has been selected, but not downloaded or run.
-- Real model traces, GGUF conversion, and app runtime wiring for the published adapter.
+- Hosted Space text runtime validation with the published GGUF. The local runtime passed, but the public Space has not been switched from mock-safe text.
+- Real text-model traces from the hosted Space.
 - Field Notes publication URL, recorded demo video URL, social post URL, and final public submission.
 
 ## Final Checks

docs/architecture-diagram.html

@@ -0,0 +1,347 @@
+<!DOCTYPE html>
+<html lang="en">
+<head>
+  <meta charset="UTF-8">
+  <meta name="viewport" content="width=device-width, initial-scale=1.0">
+  <title>Objectverse Diary Architecture Diagram</title>
+  <link href="https://fonts.googleapis.com/css2?family=JetBrains+Mono:wght@400;500;600;700&display=swap" rel="stylesheet">
+  <style>
+    * {
+      margin: 0;
+      padding: 0;
+      box-sizing: border-box;
+    }
+    
+    body {
+      font-family: 'JetBrains Mono', monospace;
+      background: #020617;
+      min-height: 100vh;
+      padding: 2rem;
+      color: white;
+    }
+    
+    .container {
+      max-width: 1200px;
+      margin: 0 auto;
+    }
+    
+    .header {
+      margin-bottom: 2rem;
+    }
+    
+    .header-row {
+      display: flex;
+      align-items: center;
+      gap: 1rem;
+      margin-bottom: 0.5rem;
+    }
+    
+    .pulse-dot {
+      width: 12px;
+      height: 12px;
+      background: #22d3ee;
+      border-radius: 50%;
+      animation: pulse 2s infinite;
+    }
+    
+    @keyframes pulse {
+      0%, 100% { opacity: 1; transform: scale(1); }
+      50% { opacity: 0.4; transform: scale(0.9); }
+    }
+    
+    h1 {
+      font-size: 1.5rem;
+      font-weight: 700;
+      letter-spacing: -0.025em;
+    }
+    
+    .subtitle {
+      color: #94a3b8;
+      font-size: 0.875rem;
+      margin-left: 1.75rem;
+    }
+    
+    .diagram-container {
+      background: rgba(15, 23, 42, 0.5);
+      border-radius: 1rem;
+      border: 1px solid #1e293b;
+      padding: 1.5rem;
+      overflow-x: auto;
+    }
+    
+    svg {
+      width: 100%;
+      min-width: 950px;
+      display: block;
+    }
+    
+    .cards {
+      display: grid;
+      grid-template-columns: repeat(auto-fit, minmax(280px, 1fr));
+      gap: 1rem;
+      margin-top: 2rem;
+    }
+    
+    .card {
+      background: rgba(15, 23, 42, 0.5);
+      border-radius: 0.75rem;
+      border: 1px solid #1e293b;
+      padding: 1.25rem;
+    }
+    
+    .card-header {
+      display: flex;
+      align-items: center;
+      gap: 0.5rem;
+      margin-bottom: 0.75rem;
+    }
+    
+    .card-dot {
+      width: 8px;
+      height: 8px;
+      border-radius: 50%;
+    }
+    
+    .card-dot.cyan { background: #22d3ee; }
+    .card-dot.emerald { background: #34d399; }
+    .card-dot.violet { background: #a78bfa; }
+    .card-dot.amber { background: #fbbf24; }
+    .card-dot.rose { background: #fb7185; }
+    
+    .card h3 {
+      font-size: 0.875rem;
+      font-weight: 600;
+    }
+    
+    .card ul {
+      list-style: none;
+      color: #94a3b8;
+      font-size: 0.75rem;
+    }
+    
+    .card li {
+      margin-bottom: 0.375rem;
+    }
+    
+    .footer {
+      text-align: center;
+      margin-top: 2rem;
+      color: #475569;
+      font-size: 0.75rem;
+    }
+  </style>
+</head>
+<body>
+  <div class="container">
+    <!-- Header -->
+    <div class="header">
+      <div class="header-row">
+        <div class="pulse-dot"></div>
+        <h1>Objectverse Diary Architecture</h1>
+      </div>
+      <p class="subtitle">Multi-layered small model pipeline for the Build Small Hackathon (An Adventure in Thousand Token Wood)</p>
+    </div>
+
+    <!-- Main Diagram -->
+    <div class="diagram-container">
+      <svg viewBox="0 0 1000 680">
+        <!-- Definitions -->
+        <defs>
+          <marker id="arrowhead" markerWidth="10" markerHeight="7" refX="9" refY="3.5" orient="auto">
+            <polygon points="0 0, 10 3.5, 0 7" fill="#64748b" />
+          </marker>
+          <pattern id="grid" width="40" height="40" patternUnits="userSpaceOnUse">
+            <path d="M 40 0 L 0 0 0 40" fill="none" stroke="#1e293b" stroke-width="0.5"/>
+          </pattern>
+        </defs>
+
+        <!-- Background Grid -->
+        <rect width="100%" height="100%" fill="url(#grid)" />
+
+        <!-- Region/Cloud Boundary (Hugging Face Space Sandbox) -->
+        <rect x="160" y="40" width="700" height="600" rx="12" fill="rgba(251, 191, 36, 0.02)" stroke="#fbbf24" stroke-width="1.2" stroke-dasharray="8,4"/>
+        <text x="175" y="62" fill="#fbbf24" font-size="10" font-weight="600">Hugging Face Space Runtime Environment</text>
+
+        <!-- Dynamic GPU Security Boundary (ZeroGPU Space Context) -->
+        <rect x="630" y="90" width="210" height="400" rx="10" fill="rgba(244, 63, 94, 0.02)" stroke="#fb7185" stroke-width="1" stroke-dasharray="4,4"/>
+        <text x="640" y="110" fill="#fb7185" font-size="8" font-weight="600">ZeroGPU Allocation Sandbox</text>
+
+        <!-- Connections (Drawn early so they layer behind components) -->
+        <!-- User -> UI Layer -->
+        <line x1="120" y1="315" x2="198" y2="315" stroke="#22d3ee" stroke-width="1.5" marker-end="url(#arrowhead)"/>
+        <text x="159" y="304" fill="#94a3b8" font-size="8" font-weight="600" text-anchor="middle">HTTPS</text>
+
+        <!-- UI Layer -> Pipeline Coordinator -->
+        <line x1="340" y1="315" x2="418" y2="315" stroke="#22d3ee" stroke-width="1.5" marker-end="url(#arrowhead)"/>
+        <text x="379" y="304" fill="#94a3b8" font-size="8" text-anchor="middle">WS/Post</text>
+
+        <!-- Coordinator -> Vision Runner -->
+        <path d="M 550 280 L 590 280 Q 610 280 610 210 L 638 210" fill="none" stroke="#34d399" stroke-width="1.5" stroke-dasharray="2,2" marker-end="url(#arrowhead)"/>
+        <text x="590" y="240" fill="#fb7185" font-size="8" text-anchor="middle">@zero_gpu</text>
+
+        <!-- Coordinator -> Llama.cpp Runner -->
+        <path d="M 550 330 L 590 330 Q 610 330 610 395 L 638 395" fill="none" stroke="#34d399" stroke-width="1.5" marker-end="url(#arrowhead)"/>
+        <text x="595" y="360" fill="#34d399" font-size="8" text-anchor="middle">Local GGUF</text>
+
+        <!-- Coordinator -> Renderer/Traces -->
+        <line x1="475" y1="365" x2="475" y2="473" stroke="#34d399" stroke-width="1.5" marker-end="url(#arrowhead)"/>
+        <text x="483" y="420" fill="#94a3b8" font-size="8">Local Save</text>
+
+        <!-- Vision Model -> HF Hub -->
+        <line x1="810" y1="210" x2="888" y2="280" stroke="#94a3b8" stroke-width="1.2" stroke-dasharray="4,4" marker-end="url(#arrowhead)"/>
+        <text x="850" y="235" fill="#94a3b8" font-size="7" text-anchor="middle">gated pull</text>
+
+        <!-- Text Model -> HF Hub -->
+        <line x1="810" y1="395" x2="888" y2="330" stroke="#94a3b8" stroke-width="1.2" stroke-dasharray="4,4" marker-end="url(#arrowhead)"/>
+        <text x="850" y="370" fill="#94a3b8" font-size="7" text-anchor="middle">GGUF pull</text>
+
+        <!-- =================================================================
+             COMPONENTS (Opaque backgrounds + Semi-transparent borders)
+             ================================================================= -->
+
+        <!-- 1. Users / Browser -->
+        <rect x="20" y="270" width="100" height="90" rx="6" fill="#0f172a" />
+        <rect x="20" y="270" width="100" height="90" rx="6" fill="rgba(30, 41, 59, 0.5)" stroke="#94a3b8" stroke-width="1.5"/>
+        <text x="70" y="295" fill="white" font-size="11" font-weight="700" text-anchor="middle">Users</text>
+        <text x="70" y="315" fill="#94a3b8" font-size="8" text-anchor="middle">Upload Image</text>
+        <text x="70" y="330" fill="#94a3b8" font-size="8" text-anchor="middle">Chat Session</text>
+        <text x="70" y="345" fill="#22d3ee" font-size="8" text-anchor="middle">Web/Mobile</text>
+
+        <!-- 2. UI Layer (Gradio) -->
+        <rect x="200" y="190" width="140" height="250" rx="8" fill="#0f172a" />
+        <rect x="200" y="190" width="140" height="250" rx="8" fill="rgba(8, 51, 68, 0.4)" stroke="#22d3ee" stroke-width="1.5"/>
+        <text x="270" y="215" fill="white" font-size="12" font-weight="700" text-anchor="middle">Gradio Web UI</text>
+        <text x="270" y="235" fill="#94a3b8" font-size="8" text-anchor="middle">app.py / src/ui/</text>
+        <line x1="210" y1="245" x2="330" y2="245" stroke="#1e293b" stroke-width="1"/>
+        <text x="215" y="265" fill="#22d3ee" font-size="8" font-weight="600">• Image Drag-Drop</text>
+        <text x="215" y="285" fill="#22d3ee" font-size="8" font-weight="600">• Persona Selector</text>
+        <text x="215" y="305" fill="#22d3ee" font-size="8" font-weight="600">• Typewriter Diary</text>
+        <text x="215" y="325" fill="#22d3ee" font-size="8" font-weight="600">• Character Chat</text>
+        <text x="215" y="345" fill="#22d3ee" font-size="8" font-weight="600">• SVG Card Render</text>
+        <line x1="210" y1="365" x2="330" y2="365" stroke="#1e293b" stroke-width="1"/>
+        <text x="270" y="385" fill="#94a3b8" font-size="8" text-anchor="middle">Example Gallery Cache</text>
+        <text x="270" y="400" fill="#22d3ee" font-size="8" text-anchor="middle">(Deterministic Baseline)</text>
+
+        <!-- 3. Pipeline Coordinator -->
+        <rect x="420" y="240" width="130" height="125" rx="6" fill="#0f172a" />
+        <rect x="420" y="240" width="130" height="125" rx="6" fill="rgba(6, 78, 59, 0.4)" stroke="#34d399" stroke-width="1.5"/>
+        <text x="485" y="265" fill="white" font-size="11" font-weight="700" text-anchor="middle">Pipeline Core</text>
+        <text x="485" y="280" fill="#94a3b8" font-size="8" text-anchor="middle">src/pipeline.py</text>
+        <line x1="430" y1="290" x2="540" y2="290" stroke="#1e293b" stroke-width="1"/>
+        <text x="435" y="305" fill="#34d399" font-size="8">• State Routing</text>
+        <text x="435" y="320" fill="#34d399" font-size="8">• Fallback Logic</text>
+        <text x="435" y="335" fill="#34d399" font-size="8">• Parse Schemas</text>
+        <text x="485" y="352" fill="#e2e8f0" font-size="7" text-anchor="middle">Pydantic validation</text>
+
+        <!-- 4. Vision Runner -->
+        <rect x="645" y="150" width="165" height="100" rx="6" fill="#0f172a" />
+        <rect x="645" y="150" width="165" height="100" rx="6" fill="rgba(120, 53, 15, 0.3)" stroke="#fbbf24" stroke-width="1.5"/>
+        <text x="727" y="175" fill="white" font-size="11" font-weight="700" text-anchor="middle">Vision Backend</text>
+        <text x="727" y="190" fill="#94a3b8" font-size="8" text-anchor="middle">src/models/vision_runner</text>
+        <line x1="655" y1="200" x2="800" y2="200" stroke="#1e293b" stroke-width="1"/>
+        <text x="660" y="215" fill="#fbbf24" font-size="8" font-weight="600">MiniCPM-V 2.6 (8B)</text>
+        <text x="660" y="230" fill="#94a3b8" font-size="8">ZeroGPU compatible</text>
+        <text x="660" y="242" fill="#fb7185" font-size="7">Fallback to mock on failure</text>
+
+        <!-- 5. Llama.cpp Runner -->
+        <rect x="645" y="335" width="165" height="110" rx="6" fill="#0f172a" />
+        <rect x="645" y="335" width="165" height="110" rx="6" fill="rgba(76, 29, 149, 0.4)" stroke="#a78bfa" stroke-width="1.5"/>
+        <text x="727" y="360" fill="white" font-size="11" font-weight="700" text-anchor="middle">Text Backend</text>
+        <text x="727" y="375" fill="#94a3b8" font-size="8" text-anchor="middle">llama_cpp_runner.py</text>
+        <line x1="655" y1="385" x2="800" y2="385" stroke="#1e293b" stroke-width="1"/>
+        <text x="660" y="400" fill="#a78bfa" font-size="8" font-weight="600">Qwen 1.5B GGUF</text>
+        <text x="660" y="415" fill="#94a3b8" font-size="8">Merged LoRA v2 Adapter</text>
+        <text x="660" y="428" fill="#fbbf24" font-size="7">Deterministic fallback runtime</text>
+
+        <!-- 6. Card Renderer & Traces (Opaque + Slate border) -->
+        <rect x="420" y="480" width="130" height="115" rx="6" fill="#0f172a" />
+        <rect x="420" y="480" width="130" height="115" rx="6" fill="rgba(30, 41, 59, 0.5)" stroke="#94a3b8" stroke-width="1.5"/>
+        <text x="485" y="505" fill="white" font-size="11" font-weight="700" text-anchor="middle">Output Services</text>
+        <text x="485" y="520" fill="#94a3b8" font-size="8" text-anchor="middle">renderer/ &amp; traces/</text>
+        <line x1="430" y1="530" x2="540" y2="530" stroke="#1e293b" stroke-width="1"/>
+        <text x="435" y="545" fill="#94a3b8" font-size="8">• Card HTML Gen</text>
+        <text x="435" y="560" fill="#94a3b8" font-size="8">• Anonymizer Traces</text>
+        <text x="435" y="575" fill="#94a3b8" font-size="8">• data/traces/*.jsonl</text>
+
+        <!-- 7. Hugging Face Hub (External Repository) -->
+        <rect x="890" y="240" width="90" height="180" rx="8" fill="#0f172a" />
+        <rect x="890" y="240" width="90" height="180" rx="8" fill="rgba(30, 41, 59, 0.5)" stroke="#94a3b8" stroke-width="1.5"/>
+        <text x="935" y="270" fill="white" font-size="11" font-weight="700" text-anchor="middle">HF Hub</text>
+        <text x="935" y="285" fill="#94a3b8" font-size="8" text-anchor="middle">Remote Assets</text>
+        <line x1="900" y1="298" x2="970" y2="298" stroke="#1e293b" stroke-width="1"/>
+        <text x="905" y="315" fill="#94a3b8" font-size="8">• SFT Dataset</text>
+        <text x="905" y="335" fill="#94a3b8" font-size="8">• LoRA Weights</text>
+        <text x="905" y="355" fill="#94a3b8" font-size="8">• GGUF Files</text>
+        <text x="905" y="375" fill="#fb7185" font-size="8">• Gate models</text>
+
+        <!-- =================================================================
+             DIAGRAM LEGEND
+             ================================================================= -->
+        <text x="180" y="515" fill="white" font-size="10" font-weight="600">Legend</text>
+        
+        <rect x="180" y="527" width="16" height="10" rx="2" fill="rgba(8, 51, 68, 0.4)" stroke="#22d3ee" stroke-width="1"/>
+        <text x="202" y="535" fill="#94a3b8" font-size="8">UI Layer (Gradio)</text>
+        
+        <rect x="180" y="543" width="16" height="10" rx="2" fill="rgba(6, 78, 59, 0.4)" stroke="#34d399" stroke-width="1"/>
+        <text x="202" y="551" fill="#94a3b8" font-size="8">Controller Layer (Python)</text>
+        
+        <rect x="180" y="559" width="16" height="10" rx="2" fill="rgba(120, 53, 15, 0.3)" stroke="#fbbf24" stroke-width="1"/>
+        <text x="202" y="567" fill="#94a3b8" font-size="8">Vision Engine (MiniCPM-V)</text>
+        
+        <rect x="180" y="575" width="16" height="10" rx="2" fill="rgba(76, 29, 149, 0.4)" stroke="#a78bfa" stroke-width="1"/>
+        <text x="202" y="583" fill="#94a3b8" font-size="8">Text Engine (Llama.cpp GGUF)</text>
+        
+        <rect x="180" y="591" width="16" height="10" rx="2" fill="rgba(30, 41, 59, 0.5)" stroke="#94a3b8" stroke-width="1"/>
+        <text x="202" y="599" fill="#94a3b8" font-size="8">External &amp; File Outputs</text>
+
+        <rect x="180" y="607" width="16" height="10" rx="2" fill="transparent" stroke="#fb7185" stroke-width="1" stroke-dasharray="2,2"/>
+        <text x="202" y="615" fill="#94a3b8" font-size="8">Security/Dynamic Hardware Group</text>
+      </svg>
+    </div>
+
+    <!-- Info Cards -->
+    <div class="cards">
+      <div class="card">
+        <div class="card-header">
+          <div class="card-dot cyan"></div>
+          <h3>UI &amp; Frontend Layer</h3>
+        </div>
+        <ul>
+          <li>• <b>English-First Copy</b>: Retro archive design, warm paper, mystery vibe</li>
+          <li>• <b>Deterministic Fallback</b>: Example gallery reads committed mock records</li>
+          <li>• <b>Interactive Sandbox</b>: Full chat session maintaining object persona</li>
+        </ul>
+      </div>
+
+      <div class="card">
+        <div class="card-header">
+          <div class="card-dot emerald"></div>
+          <h3>Pipeline Coordinator</h3>
+        </div>
+        <ul>
+          <li>• <b>Modular Routing</b>: Vision descriptions trigger first-person diaries</li>
+          <li>• <b>Pydantic Validation</b>: Strict checks on JSON output from LLM</li>
+          <li>• <b>Trace Compliance</b>: Generates anonymized session logs to JSONL files</li>
+        </ul>
+      </div>
+
+      <div class="card">
+        <div class="card-header">
+          <div class="card-dot violet"></div>
+          <h3>Dual-Engine Model Execution</h3>
+        </div>
+        <ul>
+          <li>• <b>MiniCPM-V 2.6 (8B)</b>: Runs via HF Spaces ZeroGPU dynamically</li>
+          <li>• <b>Llama.cpp (1.5B)</b>: Runs highly optimized GGUF adapter locally</li>
+          <li>• <b>Flexible Mock Fallback</b>: Ensures 100% runtime uptime for judges</li>
+        </ul>
+      </div>
+    </div>
+
+    <!-- Footer -->
+    <p class="footer">
+      Objectverse Diary • Created for Build Small Hackathon • June 2026
+    </p>
+  </div>
+</body>
+</html>

requirements-training.txt

@@ -1,2 +1,3 @@
 modal>=1,<2
 huggingface_hub>=0.34,<1
+peft>=0.19,<1

scripts/README.md

@@ -7,12 +7,15 @@ Implemented initial scripts:
 - `check_initial_stage.py`: verifies required files, runtime defaults, sample traces, pipeline, and Gradio build.
 - `generate_sample_traces.py`: creates six stable public mock traces under `data/traces/samples/`.
 - `generate_dataset.py`: creates deterministic SFT preview JSONL for schema and curation planning.
-- `prepare_curated_dataset.py`: creates 50 synthetic curated SFT rows for Modal LoRA pipeline testing.
+- `prepare_curated_dataset.py`: creates deterministic synthetic curated SFT rows; v1 defaults to 50 rows, v2 defaults to 200 rows across 40 objects and 5 modes.
 - `export_traces.py`: exports validated public sample traces to JSONL for dataset-style publishing.
 - `check_space_vlm.py`: validates MiniCPM-V object understanding on the hosted Hugging Face Space with three temporary public test images.
 - `check_llama_cpp_smoke.py`: smoke-tests the optional llama.cpp text runtime with an external GGUF model.
-- `finetune_lora.py`: validates SFT JSONL locally and defines the Modal LoRA training scaffold for the future Well-Tuned path.
+- `finetune_lora.py`: validates SFT JSONL locally and defines the Modal LoRA training scaffold with optional eval split, assistant-output-only loss, and tunable LoRA/batch settings.
+- `publish_hf_dataset.py`: validates and uploads curated JSONL files to a Hugging Face Dataset repository.
 - `publish_hf_adapter.py`: uploads a downloaded LoRA adapter folder to Hugging Face Hub.
+- `merge_lora_adapter.py`: merges a local PEFT LoRA adapter into a Hugging Face base model and saves tokenizer files.
+- `publish_hf_gguf.py`: validates and uploads a local GGUF file to a Hugging Face model repository.
 
 Expected files during implementation:
 
@@ -28,15 +31,60 @@ Modal LoRA dry-run:
   --run-name objectverse-diary-qwen15b-curated-test
 ```
 
-Modal LoRA training after explicit confirmation:
+Modal LoRA v2 dry-run for a larger curated dataset:
 
 ```bash
-modal run scripts/finetune_lora.py \
-  --dataset data/train/objectverse_sft_curated.jsonl \
-  --run-name objectverse-diary-qwen15b-curated-test \
-  --max-steps 20
+.venv/bin/python -B scripts/prepare_curated_dataset.py \
+  --version v2 \
+  --count 200 \
+  --output data/train/objectverse_sft_curated_v2.jsonl
+```
+
+Publish curated v2 dataset:
+
+```bash
+.venv/bin/python -B scripts/publish_hf_dataset.py \
+  --dataset-file data/train/objectverse_sft_curated_v2.jsonl \
+  --repo-id qqyule/objectverse-diary-sft-curated \
+  --path-in-repo objectverse_sft_curated_v2.jsonl
+```
+
+```bash
+.venv/bin/python -B scripts/finetune_lora.py \
+  --dry-run \
+  --dataset data/train/objectverse_sft_curated_v2.jsonl \
+  --run-name objectverse-diary-qwen15b-lora-v2 \
+  --max-steps 120 \
+  --learning-rate 1e-4 \
+  --max-seq-length 1536 \
+  --lora-r 32 \
+  --lora-alpha 64 \
+  --per-device-train-batch-size 2 \
+  --gradient-accumulation-steps 4 \
+  --eval-ratio 0.1 \
+  --eval-steps 20
 ```
 
+Modal LoRA v2 training:
+
+```bash
+modal run --timestamps -n objectverse-diary-qwen15b-lora-v2 scripts/finetune_lora.py \
+  --dataset data/train/objectverse_sft_curated_v2.jsonl \
+  --run-name objectverse-diary-qwen15b-lora-v2 \
+  --max-steps 120 \
+  --learning-rate 1e-4 \
+  --max-seq-length 1536 \
+  --lora-r 32 \
+  --lora-alpha 64 \
+  --per-device-train-batch-size 2 \
+  --gradient-accumulation-steps 4 \
+  --eval-ratio 0.1 \
+  --eval-steps 20
+```
+
+For epoch-based experiments, set `--max-steps 0` and provide `--num-train-epochs`.
+Assistant-output-only loss is enabled by default; pass `--no-assistant-only-loss` only for debugging full-text loss behavior.
+
 Training dependencies are intentionally separate from the Space runtime:
 
 ```bash
@@ -56,22 +104,61 @@ or configure `MODAL_TOKEN_ID` and `MODAL_TOKEN_SECRET` through your shell/secret
 After a successful Modal run, download the adapter from the output volume into ignored local exports. Modal's directory download behavior can vary; downloading individual adapter files into a directory is the safest path.
 
 ```bash
-mkdir -p exports/objectverse-diary-qwen15b-curated-test-adapter-dir
+mkdir -p exports/objectverse-diary-qwen15b-lora-v2-adapter-dir
 for file in vocab.json tokenizer_config.json tokenizer.json special_tokens_map.json merges.txt chat_template.jinja added_tokens.json adapter_model.safetensors adapter_config.json README.md; do
   modal volume get objectverse-diary-lora-output \
-    "objectverse-diary-qwen15b-curated-test/adapter/$file" \
-    "exports/objectverse-diary-qwen15b-curated-test-adapter-dir/$file"
+    "objectverse-diary-qwen15b-lora-v2/adapter/$file" \
+    "exports/objectverse-diary-qwen15b-lora-v2-adapter-dir/$file"
 done
+modal volume get objectverse-diary-lora-output \
+  objectverse-diary-qwen15b-lora-v2/metrics.json \
+  exports/objectverse-diary-qwen15b-lora-v2-adapter-dir/training_metrics.json
+modal volume get objectverse-diary-lora-output \
+  objectverse-diary-qwen15b-lora-v2/training_config.json \
+  exports/objectverse-diary-qwen15b-lora-v2-adapter-dir/training_config.json
 ```
 
 Then upload the adapter to Hugging Face Hub:
 
 ```bash
 .venv/bin/python -B scripts/publish_hf_adapter.py \
-  --adapter-dir exports/objectverse-diary-qwen15b-curated-test-adapter-dir \
-  --repo-id qqyule/objectverse-diary-qwen15b-lora
+  --adapter-dir exports/objectverse-diary-qwen15b-lora-v2-adapter-dir \
+  --repo-id qqyule/objectverse-diary-qwen15b-lora \
+  --commit-message "Upload Objectverse Diary Qwen 1.5B LoRA v2"
 ```
 
+LoRA v2 GGUF conversion and upload:
+
+```bash
+.venv/bin/python -B scripts/merge_lora_adapter.py \
+  --base-model Qwen/Qwen2.5-1.5B-Instruct \
+  --adapter exports/objectverse-diary-qwen15b-lora-v2-adapter-dir \
+  --output exports/objectverse-diary-qwen15b-lora-v2-merged-hf
+
+git clone https://github.com/ggml-org/llama.cpp.git .tmp/llama.cpp
+git -C .tmp/llama.cpp checkout 8f83d6c271d194bde2d410145a0ce73bc42e85cd
+cmake -S .tmp/llama.cpp -B .tmp/llama.cpp/build -DCMAKE_BUILD_TYPE=Release
+cmake --build .tmp/llama.cpp/build --target llama-quantize -j
+
+.venv/bin/python .tmp/llama.cpp/convert_hf_to_gguf.py \
+  exports/objectverse-diary-qwen15b-lora-v2-merged-hf \
+  --outfile models/objectverse-diary-qwen15b-lora-v2-f16.gguf \
+  --outtype f16
+
+.tmp/llama.cpp/build/bin/llama-quantize \
+  models/objectverse-diary-qwen15b-lora-v2-f16.gguf \
+  models/objectverse-diary-qwen15b-lora-v2-q4_k_m.gguf \
+  Q4_K_M
+
+.venv/bin/python -B scripts/publish_hf_gguf.py \
+  --gguf-file models/objectverse-diary-qwen15b-lora-v2-q4_k_m.gguf \
+  --repo-id qqyule/objectverse-diary-qwen15b-lora \
+  --path-in-repo objectverse-diary-qwen15b-lora-v2-q4_k_m.gguf \
+  --commit-message "Upload Objectverse Diary Qwen 1.5B LoRA v2 Q4_K_M GGUF"
+```
+
+The final Q4_K_M GGUF is ignored under `models/`. After upload, remove only generated intermediates such as the merged HF folder and F16 GGUF.
+
 Space VLM validation:
 
 ```bash
@@ -88,13 +175,13 @@ External Space changes are explicit:
 .venv/bin/python -B scripts/check_space_vlm.py --configure-space --rollback-to-mock
 ```
 
-Local GGUF smoke test after explicit confirmation:
+Local LoRA v2 GGUF smoke test:
 
 ```bash
 .venv/bin/python -B scripts/check_llama_cpp_smoke.py \
-  --model-path models/qwen2.5-1.5b-instruct-q4_k_m.gguf
+  --model-path models/objectverse-diary-qwen15b-lora-v2-q4_k_m.gguf
 ```
 
-Recommended GGUF source: `Qwen/Qwen2.5-1.5B-Instruct-GGUF`, file `qwen2.5-1.5b-instruct-q4_k_m.gguf`. Do not commit the downloaded file.
+Published GGUF source: `qqyule/objectverse-diary-qwen15b-lora`, file `objectverse-diary-qwen15b-lora-v2-q4_k_m.gguf`. Do not commit the downloaded file.
 
-Current status: mock trace generation, trace JSONL export, SFT preview generation, synthetic curated dataset publishing, optional MiniCPM-V wiring, optional llama.cpp wiring, hosted Space VLM validation tooling with non-secret probe support, local GGUF smoke helper, Modal LoRA training scaffolding, one Modal LoRA test run, and HF adapter publishing are implemented. Real model validation on Space, actual GGUF smoke, GGUF conversion, and app runtime wiring for the adapter are not completed yet.
+Current status: mock trace generation, trace JSONL export, SFT preview generation, synthetic curated v2 dataset publishing, optional MiniCPM-V wiring, optional llama.cpp wiring, hosted Space VLM validation tooling with non-secret probe support, local GGUF smoke helper, Modal LoRA training scaffolding, Modal LoRA v2 training, HF adapter publishing, GGUF conversion, GGUF upload, and local llama.cpp smoke are implemented. Real text-model validation on Space is not completed yet.

scripts/finetune_lora.py

@@ -3,7 +3,9 @@
 from __future__ import annotations
 
 import argparse
+import inspect
 import json
+import math
 import sys
 from collections.abc import Callable, Mapping, Sequence
 from dataclasses import asdict, dataclass, field
@@ -46,13 +48,48 @@ class TrainingConfig:
     run_name: str = DEFAULT_RUN_NAME
     base_model: str = DEFAULT_BASE_MODEL
     max_steps: int = 80
+    num_train_epochs: float = 3.0
     learning_rate: float = 2e-4
     max_seq_length: int = 1024
+    per_device_train_batch_size: int = 1
+    gradient_accumulation_steps: int = 4
+    eval_ratio: float = 0.1
+    eval_steps: int = 10
+    warmup_ratio: float = 0.03
+    weight_decay: float = 0.0
+    logging_steps: int = 5
+    save_total_limit: int = 2
+    seed: int = 42
+    assistant_only_loss: bool = True
     lora_r: int = 16
     lora_alpha: int = 32
     lora_dropout: float = 0.05
     target_modules: tuple[str, ...] = field(default_factory=lambda: LORA_TARGET_MODULES)
 
+    def __post_init__(self) -> None:
+        if self.max_steps < 0:
+            raise ValueError("max_steps must be 0 or greater.")
+        if self.max_steps == 0 and self.num_train_epochs <= 0:
+            raise ValueError("num_train_epochs must be greater than 0 when max_steps is 0.")
+        if self.per_device_train_batch_size < 1:
+            raise ValueError("per_device_train_batch_size must be at least 1.")
+        if self.gradient_accumulation_steps < 1:
+            raise ValueError("gradient_accumulation_steps must be at least 1.")
+        if not 0 <= self.eval_ratio < 1:
+            raise ValueError("eval_ratio must be between 0 and 1.")
+        if self.eval_steps < 1:
+            raise ValueError("eval_steps must be at least 1.")
+        if self.logging_steps < 1:
+            raise ValueError("logging_steps must be at least 1.")
+        if self.save_total_limit < 1:
+            raise ValueError("save_total_limit must be at least 1.")
+        if self.lora_r < 1:
+            raise ValueError("lora_r must be at least 1.")
+        if self.lora_alpha < 1:
+            raise ValueError("lora_alpha must be at least 1.")
+        if not 0 <= self.lora_dropout < 1:
+            raise ValueError("lora_dropout must be between 0 and 1.")
+
     def as_remote_dict(self) -> dict[str, object]:
         payload = asdict(self)
         payload["target_modules"] = list(self.target_modules)
@@ -86,11 +123,21 @@ def record_to_training_text(record: Mapping[str, object]) -> str:
     """Convert one validated chat record into a simple fallback training string."""
 
     messages = _validate_messages(record.get("messages"), line_number=None)
+    return _messages_to_training_text(messages)
+
+
+def _messages_to_training_text(
+    messages: Sequence[Mapping[str, str]],
+    *,
+    add_generation_prompt: bool = False,
+) -> str:
     blocks = []
     for message in messages:
         role = str(message["role"]).strip().lower()
         content = str(message["content"]).strip()
         blocks.append(f"{role}:\n{content}")
+    if add_generation_prompt:
+        blocks.append("assistant:\n")
     return "\n\n".join(blocks).strip()
 
 
@@ -145,15 +192,29 @@ def _dry_run_summary(
     config: TrainingConfig,
 ) -> dict[str, object]:
     first_text = record_to_training_text(records[0])
+    eval_count = _eval_record_count(len(records), config.eval_ratio)
     return {
         "mode": "dry-run",
         "dataset": str(dataset),
         "record_count": len(records),
+        "train_record_count": len(records) - eval_count,
+        "eval_record_count": eval_count,
         "base_model": config.base_model,
         "run_name": config.run_name,
         "max_steps": config.max_steps,
+        "num_train_epochs": config.num_train_epochs,
         "learning_rate": config.learning_rate,
         "max_seq_length": config.max_seq_length,
+        "per_device_train_batch_size": config.per_device_train_batch_size,
+        "gradient_accumulation_steps": config.gradient_accumulation_steps,
+        "effective_batch_size": (
+            config.per_device_train_batch_size * config.gradient_accumulation_steps
+        ),
+        "eval_ratio": config.eval_ratio,
+        "eval_steps": config.eval_steps,
+        "warmup_ratio": config.warmup_ratio,
+        "weight_decay": config.weight_decay,
+        "assistant_only_loss": config.assistant_only_loss,
         "lora": {
             "r": config.lora_r,
             "alpha": config.lora_alpha,
@@ -184,7 +245,6 @@ def _train_lora_impl(
     from transformers import (
         AutoModelForCausalLM,
         AutoTokenizer,
-        DataCollatorForLanguageModeling,
         Trainer,
         TrainingArguments,
     )
@@ -218,49 +278,46 @@ def _train_lora_impl(
     model.print_trainable_parameters()
 
     dataset = Dataset.from_list(
-        [{"text": _format_training_text(record, tokenizer)} for record in records]
+        [
+            _tokenize_training_example(
+                record,
+                tokenizer,
+                max_length=config.max_seq_length,
+                assistant_only_loss=config.assistant_only_loss,
+            )
+            for record in records
+        ]
     )
+    train_dataset, eval_dataset = _split_dataset(dataset, config)
 
-    def tokenize_batch(batch: Mapping[str, list[str]]) -> dict[str, object]:
-        return tokenizer(
-            batch["text"],
-            truncation=True,
-            max_length=config.max_seq_length,
-            padding=False,
-        )
-
-    tokenized = dataset.map(
-        tokenize_batch,
-        batched=True,
-        remove_columns=["text"],
-        desc="Tokenize Objectverse Diary SFT records",
+    training_kwargs = _training_arguments_kwargs(
+        output_dir=output_path / "trainer",
+        config=config,
+        has_eval=eval_dataset is not None,
+        training_arguments_cls=TrainingArguments,
     )
+    training_kwargs["fp16"] = torch.cuda.is_available()
 
-    training_args = TrainingArguments(
-        output_dir=str(output_path / "trainer"),
-        max_steps=config.max_steps,
-        per_device_train_batch_size=1,
-        gradient_accumulation_steps=4,
-        learning_rate=config.learning_rate,
-        logging_steps=5,
-        save_strategy="no",
-        fp16=torch.cuda.is_available(),
-        report_to=[],
-        optim="adamw_torch",
-    )
+    training_args = TrainingArguments(**training_kwargs)
     trainer = Trainer(
         model=model,
         args=training_args,
-        train_dataset=tokenized,
-        data_collator=DataCollatorForLanguageModeling(tokenizer=tokenizer, mlm=False),
+        train_dataset=train_dataset,
+        eval_dataset=eval_dataset,
+        data_collator=_build_supervised_data_collator(tokenizer, torch),
     )
     train_result = trainer.train()
+    eval_metrics: dict[str, object] = {}
+    if eval_dataset is not None:
+        eval_metrics = dict(trainer.evaluate())
 
     model.save_pretrained(adapter_path)
     tokenizer.save_pretrained(adapter_path)
 
     metrics = dict(train_result.metrics)
-    metrics["train_records"] = len(records)
+    metrics.update(eval_metrics)
+    metrics["train_records"] = len(train_dataset)
+    metrics["eval_records"] = len(eval_dataset) if eval_dataset is not None else 0
     metrics["base_model"] = config.base_model
     (output_path / "metrics.json").write_text(
         json.dumps(metrics, indent=2, sort_keys=True),
@@ -278,11 +335,157 @@ def _train_lora_impl(
         "mode": "modal-training",
         "run_name": config.run_name,
         "record_count": len(records),
+        "train_record_count": len(train_dataset),
+        "eval_record_count": len(eval_dataset) if eval_dataset is not None else 0,
         "adapter_path": str(adapter_path),
         "metrics_path": str(output_path / "metrics.json"),
     }
 
 
+def _tokenize_training_example(
+    record: Mapping[str, object],
+    tokenizer: Any,
+    *,
+    max_length: int,
+    assistant_only_loss: bool,
+) -> dict[str, list[int]]:
+    full_text = _format_training_text(record, tokenizer)
+    encoded = tokenizer(
+        full_text,
+        truncation=True,
+        max_length=max_length,
+        padding=False,
+        add_special_tokens=False,
+    )
+    input_ids = list(encoded["input_ids"])
+    labels = list(input_ids)
+
+    if assistant_only_loss:
+        prompt_text = _format_prompt_text(record, tokenizer)
+        prompt_encoded = tokenizer(
+            prompt_text,
+            truncation=True,
+            max_length=max_length,
+            padding=False,
+            add_special_tokens=False,
+        )
+        mask_count = min(len(prompt_encoded["input_ids"]), len(labels))
+        labels[:mask_count] = [-100] * mask_count
+        if not any(label != -100 for label in labels):
+            raise ValueError(
+                "max_seq_length truncates all assistant labels; increase max_seq_length."
+            )
+
+    return {
+        "input_ids": input_ids,
+        "attention_mask": list(encoded["attention_mask"]),
+        "labels": labels,
+    }
+
+
+def _split_dataset(dataset: Any, config: TrainingConfig) -> tuple[Any, Any | None]:
+    eval_count = _eval_record_count(len(dataset), config.eval_ratio)
+    if eval_count == 0:
+        return dataset, None
+    split = dataset.train_test_split(test_size=eval_count, shuffle=True, seed=config.seed)
+    return split["train"], split["test"]
+
+
+def _eval_record_count(record_count: int, eval_ratio: float) -> int:
+    if record_count < 2 or eval_ratio <= 0:
+        return 0
+    return max(1, min(record_count - 1, math.ceil(record_count * eval_ratio)))
+
+
+def _training_arguments_kwargs(
+    *,
+    output_dir: Path,
+    config: TrainingConfig,
+    has_eval: bool,
+    training_arguments_cls: Any | None = None,
+) -> dict[str, object]:
+    kwargs: dict[str, object] = {
+        "output_dir": str(output_dir),
+        "per_device_train_batch_size": config.per_device_train_batch_size,
+        "gradient_accumulation_steps": config.gradient_accumulation_steps,
+        "learning_rate": config.learning_rate,
+        "logging_steps": config.logging_steps,
+        "warmup_ratio": config.warmup_ratio,
+        "weight_decay": config.weight_decay,
+        "report_to": [],
+        "optim": "adamw_torch",
+        "seed": config.seed,
+        "data_seed": config.seed,
+    }
+    if config.max_steps > 0:
+        kwargs["max_steps"] = config.max_steps
+    else:
+        kwargs["num_train_epochs"] = config.num_train_epochs
+
+    if has_eval:
+        kwargs.update(
+            {
+                "eval_steps": config.eval_steps,
+                "save_steps": config.eval_steps,
+                "save_strategy": "steps",
+                "save_total_limit": config.save_total_limit,
+                "load_best_model_at_end": True,
+                "metric_for_best_model": "eval_loss",
+                "greater_is_better": False,
+            }
+        )
+        if training_arguments_cls is None:
+            kwargs["eval_strategy"] = "steps"
+        else:
+            _set_eval_strategy_kwarg(kwargs, training_arguments_cls, "steps")
+    else:
+        kwargs["save_strategy"] = "no"
+
+    return kwargs
+
+
+def _set_eval_strategy_kwarg(
+    kwargs: dict[str, object],
+    training_arguments_cls: Any,
+    strategy: str,
+) -> None:
+    parameters = inspect.signature(training_arguments_cls.__init__).parameters
+    if "eval_strategy" in parameters:
+        kwargs["eval_strategy"] = strategy
+    elif "evaluation_strategy" in parameters:
+        kwargs["evaluation_strategy"] = strategy
+    else:
+        kwargs["do_eval"] = strategy != "no"
+
+
+def _build_supervised_data_collator(tokenizer: Any, torch_module: Any) -> Callable:
+    def collate(features: list[Mapping[str, list[int]]]) -> dict[str, object]:
+        labels = [list(feature["labels"]) for feature in features]
+        model_features = [
+            {
+                "input_ids": list(feature["input_ids"]),
+                "attention_mask": list(feature["attention_mask"]),
+            }
+            for feature in features
+        ]
+        batch = tokenizer.pad(model_features, padding=True, return_tensors="pt")
+        max_length = batch["input_ids"].shape[1]
+        label_tensor = torch_module.full(
+            (len(labels), max_length),
+            -100,
+            dtype=torch_module.long,
+        )
+        for index, label in enumerate(labels):
+            label_tensor[index, : len(label)] = torch_module.tensor(
+                label,
+                dtype=torch_module.long,
+            )
+        batch["labels"] = label_tensor
+        return batch
+
+    return collate
+
+
 def _training_config_from_payload(payload: Mapping[str, object]) -> TrainingConfig:
     target_modules = payload.get("target_modules", LORA_TARGET_MODULES)
     if not isinstance(target_modules, Sequence) or isinstance(target_modules, (str, bytes)):
@@ -291,8 +494,19 @@ def _training_config_from_payload(payload: Mapping[str, object]) -> TrainingConf
         run_name=str(payload.get("run_name", DEFAULT_RUN_NAME)),
         base_model=str(payload.get("base_model", DEFAULT_BASE_MODEL)),
         max_steps=int(payload.get("max_steps", 80)),
+        num_train_epochs=float(payload.get("num_train_epochs", 3.0)),
         learning_rate=float(payload.get("learning_rate", 2e-4)),
         max_seq_length=int(payload.get("max_seq_length", 1024)),
+        per_device_train_batch_size=int(payload.get("per_device_train_batch_size", 1)),
+        gradient_accumulation_steps=int(payload.get("gradient_accumulation_steps", 4)),
+        eval_ratio=float(payload.get("eval_ratio", 0.1)),
+        eval_steps=int(payload.get("eval_steps", 10)),
+        warmup_ratio=float(payload.get("warmup_ratio", 0.03)),
+        weight_decay=float(payload.get("weight_decay", 0.0)),
+        logging_steps=int(payload.get("logging_steps", 5)),
+        save_total_limit=int(payload.get("save_total_limit", 2)),
+        seed=int(payload.get("seed", 42)),
+        assistant_only_loss=bool(payload.get("assistant_only_loss", True)),
         lora_r=int(payload.get("lora_r", 16)),
         lora_alpha=int(payload.get("lora_alpha", 32)),
         lora_dropout=float(payload.get("lora_dropout", 0.05)),
@@ -302,16 +516,39 @@ def _training_config_from_payload(payload: Mapping[str, object]) -> TrainingConf
 
 def _format_training_text(record: Mapping[str, object], tokenizer: Any) -> str:
     messages = _validate_messages(record.get("messages"), line_number=None)
+    return _format_messages(messages, tokenizer, add_generation_prompt=False)
+
+
+def _format_prompt_text(record: Mapping[str, object], tokenizer: Any) -> str:
+    messages = _validate_messages(record.get("messages"), line_number=None)
+    assistant_indices = [
+        index for index, message in enumerate(messages) if message["role"].lower() == "assistant"
+    ]
+    if not assistant_indices:
+        raise ValueError("assistant_only_loss requires at least one assistant message.")
+    prompt_messages = messages[: assistant_indices[-1]]
+    return _format_messages(prompt_messages, tokenizer, add_generation_prompt=True)
+
+
+def _format_messages(
+    messages: Sequence[Mapping[str, str]],
+    tokenizer: Any,
+    *,
+    add_generation_prompt: bool,
+) -> str:
     if hasattr(tokenizer, "apply_chat_template"):
         try:
             return tokenizer.apply_chat_template(
                 messages,
                 tokenize=False,
-                add_generation_prompt=False,
+                add_generation_prompt=add_generation_prompt,
             )
         except Exception:
             pass
-    return record_to_training_text(record)
+    return _messages_to_training_text(
+        messages,
+        add_generation_prompt=add_generation_prompt,
+    )
 
 
 def _print_json(payload: Mapping[str, object]) -> None:
@@ -323,8 +560,22 @@ def _build_config_from_args(args: argparse.Namespace) -> TrainingConfig:
         run_name=args.run_name,
         base_model=args.base_model,
         max_steps=args.max_steps,
+        num_train_epochs=args.num_train_epochs,
         learning_rate=args.learning_rate,
         max_seq_length=args.max_seq_length,
+        per_device_train_batch_size=args.per_device_train_batch_size,
+        gradient_accumulation_steps=args.gradient_accumulation_steps,
+        eval_ratio=args.eval_ratio,
+        eval_steps=args.eval_steps,
+        warmup_ratio=args.warmup_ratio,
+        weight_decay=args.weight_decay,
+        logging_steps=args.logging_steps,
+        save_total_limit=args.save_total_limit,
+        seed=args.seed,
+        assistant_only_loss=args.assistant_only_loss,
+        lora_r=args.lora_r,
+        lora_alpha=args.lora_alpha,
+        lora_dropout=args.lora_dropout,
     )
 
 
@@ -334,8 +585,22 @@ def _parse_args(argv: Sequence[str] | None = None) -> argparse.Namespace:
     parser.add_argument("--run-name", default=DEFAULT_RUN_NAME)
     parser.add_argument("--base-model", default=DEFAULT_BASE_MODEL)
     parser.add_argument("--max-steps", type=int, default=80)
+    parser.add_argument("--num-train-epochs", type=float, default=3.0)
     parser.add_argument("--learning-rate", type=float, default=2e-4)
     parser.add_argument("--max-seq-length", type=int, default=1024)
+    parser.add_argument("--per-device-train-batch-size", type=int, default=1)
+    parser.add_argument("--gradient-accumulation-steps", type=int, default=4)
+    parser.add_argument("--eval-ratio", type=float, default=0.1)
+    parser.add_argument("--eval-steps", type=int, default=10)
+    parser.add_argument("--warmup-ratio", type=float, default=0.03)
+    parser.add_argument("--weight-decay", type=float, default=0.0)
+    parser.add_argument("--logging-steps", type=int, default=5)
+    parser.add_argument("--save-total-limit", type=int, default=2)
+    parser.add_argument("--seed", type=int, default=42)
+    parser.add_argument("--assistant-only-loss", action=argparse.BooleanOptionalAction, default=True)
+    parser.add_argument("--lora-r", type=int, default=16)
+    parser.add_argument("--lora-alpha", type=int, default=32)
+    parser.add_argument("--lora-dropout", type=float, default=0.05)
     parser.add_argument("--dry-run", action="store_true")
     return parser.parse_args(argv)
 
@@ -390,8 +655,22 @@ if modal is not None:
         run_name: str = DEFAULT_RUN_NAME,
         base_model: str = DEFAULT_BASE_MODEL,
         max_steps: int = 80,
+        num_train_epochs: float = 3.0,
         learning_rate: float = 2e-4,
         max_seq_length: int = 1024,
+        per_device_train_batch_size: int = 1,
+        gradient_accumulation_steps: int = 4,
+        eval_ratio: float = 0.1,
+        eval_steps: int = 10,
+        warmup_ratio: float = 0.03,
+        weight_decay: float = 0.0,
+        logging_steps: int = 5,
+        save_total_limit: int = 2,
+        seed: int = 42,
+        assistant_only_loss: bool = True,
+        lora_r: int = 16,
+        lora_alpha: int = 32,
+        lora_dropout: float = 0.05,
         dry_run: bool = False,
     ) -> None:
         result = run_training_entrypoint(
@@ -400,8 +679,22 @@ if modal is not None:
                 run_name=run_name,
                 base_model=base_model,
                 max_steps=max_steps,
+                num_train_epochs=num_train_epochs,
                 learning_rate=learning_rate,
                 max_seq_length=max_seq_length,
+                per_device_train_batch_size=per_device_train_batch_size,
+                gradient_accumulation_steps=gradient_accumulation_steps,
+                eval_ratio=eval_ratio,
+                eval_steps=eval_steps,
+                warmup_ratio=warmup_ratio,
+                weight_decay=weight_decay,
+                logging_steps=logging_steps,
+                save_total_limit=save_total_limit,
+                seed=seed,
+                assistant_only_loss=assistant_only_loss,
+                lora_r=lora_r,
+                lora_alpha=lora_alpha,
+                lora_dropout=lora_dropout,
             ),
             dry_run=dry_run,
             allow_remote=True,

scripts/merge_lora_adapter.py

@@ -0,0 +1,155 @@
+"""Merge an Objectverse Diary LoRA adapter into its base Hugging Face model."""
+
+from __future__ import annotations
+
+import argparse
+import json
+from pathlib import Path
+from typing import Any
+
+
+ADAPTER_WEIGHT_FILES = ("adapter_model.safetensors", "adapter_model.bin")
+
+
+def validate_adapter_source(adapter: str | Path, *, base_model: str) -> dict[str, object]:
+    adapter_text = str(adapter)
+    adapter_path = Path(adapter_text)
+    if adapter_path.exists():
+        if not adapter_path.is_dir():
+            raise ValueError(f"Adapter path is not a directory: {adapter_path}")
+        config_path = adapter_path / "adapter_config.json"
+        if not config_path.exists():
+            raise ValueError(f"Adapter directory is missing adapter_config.json: {adapter_path}")
+        if not any((adapter_path / name).exists() for name in ADAPTER_WEIGHT_FILES):
+            raise ValueError(
+                "Adapter directory is missing adapter_model.safetensors or adapter_model.bin."
+            )
+        config = _read_adapter_config(config_path)
+        configured_base = config.get("base_model_name_or_path")
+        if configured_base and str(configured_base) != base_model:
+            raise ValueError(
+                f"Adapter base model is {configured_base!r}, expected {base_model!r}."
+            )
+        return {
+            "adapter": str(adapter_path),
+            "adapter_type": "local",
+            "adapter_base_model": configured_base or "",
+        }
+
+    if "/" not in adapter_text:
+        raise FileNotFoundError(f"Adapter source does not exist: {adapter_text}")
+    return {
+        "adapter": adapter_text,
+        "adapter_type": "hub",
+        "adapter_base_model": "",
+    }
+
+
+def plan_merge(
+    *,
+    base_model: str,
+    adapter: str | Path,
+    output: Path,
+    dry_run: bool,
+) -> dict[str, object]:
+    summary = validate_adapter_source(adapter, base_model=base_model)
+    summary.update(
+        {
+            "base_model": base_model,
+            "output": str(output),
+            "dry_run": dry_run,
+        }
+    )
+    if dry_run:
+        summary["merged"] = False
+        return summary
+
+    merge_lora_adapter(
+        base_model=base_model,
+        adapter=str(adapter),
+        output=output,
+    )
+    summary["merged"] = True
+    summary["files"] = sorted(path.name for path in output.iterdir() if path.is_file())
+    return summary
+
+
+def merge_lora_adapter(
+    *,
+    base_model: str,
+    adapter: str,
+    output: Path,
+) -> None:
+    from peft import PeftModel
+    from transformers import AutoModelForCausalLM, AutoTokenizer
+
+    output.mkdir(parents=True, exist_ok=True)
+    model = AutoModelForCausalLM.from_pretrained(
+        base_model,
+        torch_dtype="auto",
+        device_map={"": "cpu"},
+        low_cpu_mem_usage=True,
+    )
+    peft_model = PeftModel.from_pretrained(model, adapter)
+    merged = peft_model.merge_and_unload(safe_merge=True)
+    merged.save_pretrained(
+        output,
+        safe_serialization=True,
+        max_shard_size="2GB",
+    )
+
+    tokenizer = AutoTokenizer.from_pretrained(adapter if Path(adapter).exists() else base_model)
+    tokenizer.save_pretrained(output)
+
+    metadata = {
+        "base_model": base_model,
+        "adapter": adapter,
+        "output": str(output),
+        "format": "merged-hf",
+    }
+    (output / "objectverse_merge_metadata.json").write_text(
+        json.dumps(metadata, indent=2, sort_keys=True),
+        encoding="utf-8",
+    )
+
+
+def _read_adapter_config(config_path: Path) -> dict[str, object]:
+    try:
+        payload = json.loads(config_path.read_text(encoding="utf-8"))
+    except json.JSONDecodeError as exc:
+        raise ValueError(f"Invalid adapter_config.json: {exc.msg}") from exc
+    if not isinstance(payload, dict):
+        raise ValueError("adapter_config.json must contain a JSON object.")
+    return payload
+
+
+def _print_json(payload: dict[str, Any]) -> None:
+    print(json.dumps(payload, indent=2, sort_keys=True), flush=True)
+
+
+def _parse_args() -> argparse.Namespace:
+    parser = argparse.ArgumentParser(description=__doc__)
+    parser.add_argument("--base-model", required=True)
+    parser.add_argument("--adapter", required=True)
+    parser.add_argument("--output", type=Path, required=True)
+    parser.add_argument("--dry-run", action="store_true")
+    return parser.parse_args()
+
+
+def main() -> None:
+    args = _parse_args()
+    _print_json(
+        plan_merge(
+            base_model=args.base_model,
+            adapter=args.adapter,
+            output=args.output,
+            dry_run=args.dry_run,
+        )
+    )
+
+
+if __name__ == "__main__":
+    try:
+        main()
+    except Exception as exc:
+        raise SystemExit(str(exc)) from exc

scripts/prepare_curated_dataset.py

@@ -16,8 +16,11 @@ from src.models.schema import DiaryEntry, ObjectInfo, ObjectUnderstanding, Perso
 
 
 DEFAULT_OUTPUT_PATH = Path("data/train/objectverse_sft_curated.jsonl")
+DEFAULT_V2_OUTPUT_PATH = Path("data/train/objectverse_sft_curated_v2.jsonl")
 DEFAULT_COUNT = 50
-SOURCE = "objectverse-diary-synthetic-curated-v1"
+DEFAULT_V2_COUNT = 200
+SOURCE_V1 = "objectverse-diary-synthetic-curated-v1"
+SOURCE_V2 = "objectverse-diary-synthetic-curated-v2"
 
 SYSTEM_PROMPT = (
     "You are Objectverse Diary, an English-first small-model assistant. "
@@ -91,6 +94,226 @@ OBJECTS = [
     },
 ]
 
+OBJECTS_V2 = [
+    *(
+        dict(
+            obj,
+            scene_detail=f"resting in the {obj['context']} with a history no one inventoried",
+        )
+        for obj in OBJECTS
+    ),
+    {
+        "name": "wireless earbud case",
+        "features": ["rounded white shell", "tiny hinge", "charging light"],
+        "context": "commuter bag",
+        "memory": "held two small arguments against silence through a crowded train",
+        "scene_detail": "buried beside lint, receipts, and one forgotten mint",
+    },
+    {
+        "name": "transit card",
+        "features": ["scuffed plastic", "faded corner", "thin blue stripe"],
+        "context": "wallet slot",
+        "memory": "opened gates for mornings that were already late",
+        "scene_detail": "pressed flat under coins and expired coupons",
+    },
+    {
+        "name": "canvas tote bag",
+        "features": ["creased cotton", "ink logo", "soft handles"],
+        "context": "entryway floor",
+        "memory": "carried groceries, books, and ambitions heavier than both",
+        "scene_detail": "slumped open with a receipt still clinging inside",
+    },
+    {
+        "name": "cracked phone case",
+        "features": ["clear plastic", "corner crack", "fingerprint haze"],
+        "context": "bedside table",
+        "memory": "took the impact so the glowing rectangle could remain innocent",
+        "scene_detail": "lying face down after another nervous scroll",
+    },
+    {
+        "name": "lip balm tube",
+        "features": ["twisted cap", "pocket scratches", "worn label"],
+        "context": "coat pocket",
+        "memory": "answered every small weather emergency without being thanked",
+        "scene_detail": "rolling between keys and a folded train ticket",
+    },
+    {
+        "name": "medicine organizer",
+        "features": ["clear lids", "weekday letters", "plastic hinges"],
+        "context": "bathroom shelf",
+        "memory": "sorted tiny promises into seven obedient compartments",
+        "scene_detail": "waiting under fluorescent light with Monday already open",
+    },
+    {
+        "name": "travel toothbrush",
+        "features": ["folding handle", "blue bristles", "vented cap"],
+        "context": "hotel sink",
+        "memory": "kept a mouth honest in rooms that forgot every guest",
+        "scene_detail": "balanced near a wrapped soap and a paper cup",
+    },
+    {
+        "name": "passport cover",
+        "features": ["navy leather", "creased spine", "stitched edge"],
+        "context": "carry-on pocket",
+        "memory": "guarded borders, delays, and a face trying to look awake",
+        "scene_detail": "wedged beside boarding papers and a silent pen",
+    },
+    {
+        "name": "boarding pass stub",
+        "features": ["thermal paper", "torn edge", "gate code"],
+        "context": "jacket pocket",
+        "memory": "proved a journey happened after the airport swallowed the day",
+        "scene_detail": "softened by rain and folded into four tired rectangles",
+    },
+    {
+        "name": "hotel keycard",
+        "features": ["matte plastic", "blank stripe", "room-number sleeve"],
+        "context": "nightstand",
+        "memory": "opened a temporary room for a temporary version of its human",
+        "scene_detail": "resting beside a glass of water no one trusted",
+    },
+    {
+        "name": "remote control",
+        "features": ["rubber buttons", "battery door scar", "dusty edges"],
+        "context": "sofa cushion",
+        "memory": "changed channels while nobody changed their mind",
+        "scene_detail": "half-sunk between cushions with one crumb for company",
+    },
+    {
+        "name": "reading glasses",
+        "features": ["thin frames", "smudged lenses", "bent temple"],
+        "context": "book stack",
+        "memory": "made small letters confess their meaning at midnight",
+        "scene_detail": "left open across a page that was never finished",
+    },
+    {
+        "name": "glasses case",
+        "features": ["hard shell", "soft lining", "snap hinge"],
+        "context": "desk drawer",
+        "memory": "protected fragile clarity from the tyranny of keys",
+        "scene_detail": "waiting in darkness with a paperclip pressed to its side",
+    },
+    {
+        "name": "wristwatch",
+        "features": ["scratched face", "brown strap", "small crown"],
+        "context": "dresser tray",
+        "memory": "measured days while humans pretended not to be measured",
+        "scene_detail": "stopped beside coins and a single loose button",
+    },
+    {
+        "name": "hair clip",
+        "features": ["amber plastic", "tiny teeth", "curved spring"],
+        "context": "bathroom counter",
+        "memory": "held chaos together for meetings, errands, and almost-crying",
+        "scene_detail": "resting near a fogged mirror and stray strands",
+    },
+    {
+        "name": "laundry token",
+        "features": ["round brass", "machine number", "dulled rim"],
+        "context": "laundry room",
+        "memory": "bought one more spin for clothes that knew too much",
+        "scene_detail": "cool in a palm smelling faintly of detergent",
+    },
+    {
+        "name": "refrigerator magnet",
+        "features": ["painted souvenir", "flat magnet back", "chipped corner"],
+        "context": "kitchen door",
+        "memory": "held reminders in place while everyone forgot the reason",
+        "scene_detail": "pinning a grocery list under a blue-white hum",
+    },
+    {
+        "name": "grocery receipt",
+        "features": ["curled paper", "faded ink", "long total"],
+        "context": "kitchen counter",
+        "memory": "itemized hunger, soap, and one unnecessary chocolate bar",
+        "scene_detail": "curling beside fruit that ripened too quickly",
+    },
+    {
+        "name": "spice jar",
+        "features": ["glass body", "red powder", "metal lid"],
+        "context": "kitchen shelf",
+        "memory": "made bland evenings briefly remember a warmer country",
+        "scene_detail": "standing in a row of louder labels",
+    },
+    {
+        "name": "cutting board",
+        "features": ["wood grain", "knife marks", "rounded corner"],
+        "context": "kitchen island",
+        "memory": "received every chopped plan without flinching",
+        "scene_detail": "drying upright after a meal nobody photographed",
+    },
+    {
+        "name": "ceramic bowl",
+        "features": ["blue rim", "tiny chip", "glazed curve"],
+        "context": "dish rack",
+        "memory": "held soup, cereal, and one quiet apology",
+        "scene_detail": "tilted beside plates still warm from rinse water",
+    },
+    {
+        "name": "reusable chopsticks",
+        "features": ["dark bamboo", "tapered tips", "cloth sleeve"],
+        "context": "lunch bag",
+        "memory": "lifted noodles through ordinary hunger and office gossip",
+        "scene_detail": "tucked into a sleeve with a soy sauce stain",
+    },
+    {
+        "name": "tea tin",
+        "features": ["green metal", "tight lid", "leaf dust"],
+        "context": "pantry shelf",
+        "memory": "kept rain-colored leaves ready for small recoveries",
+        "scene_detail": "quiet behind cereal boxes and a jar of almonds",
+    },
+    {
+        "name": "sticky note stack",
+        "features": ["yellow pages", "curled edge", "faint adhesive"],
+        "context": "monitor base",
+        "memory": "accepted urgent thoughts that became decorative fossils",
+        "scene_detail": "leaning under a monitor's cold rectangular sun",
+    },
+    {
+        "name": "binder clip",
+        "features": ["black steel", "silver arms", "pinched mouth"],
+        "context": "paper tray",
+        "memory": "held loose pages together when ideas tried to scatter",
+        "scene_detail": "biting a stack marked later in blue ink",
+    },
+    {
+        "name": "fountain pen",
+        "features": ["black barrel", "gold nib", "ink stain"],
+        "context": "notebook margin",
+        "memory": "turned hesitation into lines that looked deliberate",
+        "scene_detail": "uncapped beside a sentence crossed out twice",
+    },
+    {
+        "name": "old ticket stub",
+        "features": ["creased paper", "seat number", "torn perforation"],
+        "context": "memory box",
+        "memory": "survived the event after the applause became dust",
+        "scene_detail": "pressed under postcards and a dried ribbon",
+    },
+    {
+        "name": "candle jar",
+        "features": ["smoked glass", "wax tunnel", "blackened wick"],
+        "context": "window ledge",
+        "memory": "made one room pretend to be softer than it was",
+        "scene_detail": "cooled beside a window with rain on the other side",
+    },
+    {
+        "name": "alarm clock",
+        "features": ["round face", "plastic feet", "stubborn button"],
+        "context": "bedside shelf",
+        "memory": "tore people from dreams and was hated for being correct",
+        "scene_detail": "facing a bed that negotiated with every morning",
+    },
+    {
+        "name": "tape measure",
+        "features": ["yellow tape", "lock switch", "metal hook"],
+        "context": "tool drawer",
+        "memory": "proved shelves, windows, and ambitions were smaller than claimed",
+        "scene_detail": "coiled beside screws and one pencil shaved short",
+    },
+]
+
 MODE_PROFILES = {
     "Cynical": {
         "mood": "tired but sharply observant",
@@ -125,15 +348,24 @@ MODE_PROFILES = {
 }
 
 
-def build_curated_records(count: int = DEFAULT_COUNT) -> list[dict[str, object]]:
+def build_curated_records(
+    count: int | None = None,
+    *,
+    version: str = "v1",
+) -> list[dict[str, object]]:
+    version = _validate_version(version)
+    if count is None:
+        count = DEFAULT_V2_COUNT if version == "v2" else DEFAULT_COUNT
     if count < 1:
         raise ValueError("count must be at least 1")
 
+    objects = _objects_for_version(version)
+    source = _source_for_version(version)
     records: list[dict[str, object]] = []
     for index in range(count):
-        obj = OBJECTS[index % len(OBJECTS)]
-        mode = MODES[(index // len(OBJECTS)) % len(MODES)]
-        record_id = f"curated-synthetic-{index + 1:04d}"
+        obj = objects[index % len(objects)]
+        mode = MODES[(index // len(objects)) % len(MODES)]
+        record_id = _record_id(version, index)
         understanding = _build_object_understanding(obj)
         persona = _build_persona(obj, mode)
         diary = _build_diary(obj, mode, persona.persona, index)
@@ -141,31 +373,29 @@ def build_curated_records(count: int = DEFAULT_COUNT) -> list[dict[str, object]]
             "persona": persona.persona.model_dump(mode="json"),
             "diary": diary.model_dump(mode="json"),
         }
-        records.append(
-            {
-                "id": record_id,
-                "source": SOURCE,
-                "split": "train",
-                "mode": mode,
-                "object_description": _object_description(obj),
-                "object_understanding": understanding.model_dump(mode="json"),
-                "curation_notes": (
-                    "Synthetic curated row: no private photo, no personal identifier, "
-                    "English-first output with Chinese helper text."
-                ),
-                "messages": [
-                    {"role": "system", "content": SYSTEM_PROMPT},
-                    {
-                        "role": "user",
-                        "content": _user_prompt(understanding.model_dump(mode="json"), mode),
-                    },
-                    {
-                        "role": "assistant",
-                        "content": json.dumps(assistant_payload, ensure_ascii=False),
-                    },
-                ],
-            }
-        )
+        record = {
+            "id": record_id,
+            "source": source,
+            "split": "train",
+            "mode": mode,
+            "object_description": _object_description(obj),
+            "object_understanding": understanding.model_dump(mode="json"),
+            "curation_notes": _curation_notes(version),
+            "messages": [
+                {"role": "system", "content": SYSTEM_PROMPT},
+                {
+                    "role": "user",
+                    "content": _user_prompt(understanding.model_dump(mode="json"), mode),
+                },
+                {
+                    "role": "assistant",
+                    "content": json.dumps(assistant_payload, ensure_ascii=False),
+                },
+            ],
+        }
+        if version == "v2":
+            record["scene_detail"] = str(obj["scene_detail"])
+        records.append(record)
     return records
 
 
@@ -176,8 +406,48 @@ def write_jsonl(records: Sequence[Mapping[str, object]], output_path: Path) -> P
     return output_path
 
 
-def prepare_curated_dataset(output_path: Path = DEFAULT_OUTPUT_PATH, count: int = DEFAULT_COUNT) -> Path:
-    return write_jsonl(build_curated_records(count), output_path)
+def prepare_curated_dataset(
+    output_path: Path | None = None,
+    count: int | None = None,
+    *,
+    version: str = "v1",
+) -> Path:
+    version = _validate_version(version)
+    if output_path is None:
+        output_path = DEFAULT_V2_OUTPUT_PATH if version == "v2" else DEFAULT_OUTPUT_PATH
+    return write_jsonl(build_curated_records(count, version=version), output_path)
+
+
+def _validate_version(version: str) -> str:
+    if version not in {"v1", "v2"}:
+        raise ValueError("version must be 'v1' or 'v2'.")
+    return version
+
+
+def _objects_for_version(version: str) -> Sequence[Mapping[str, object]]:
+    return OBJECTS_V2 if version == "v2" else OBJECTS
+
+
+def _source_for_version(version: str) -> str:
+    return SOURCE_V2 if version == "v2" else SOURCE_V1
+
+
+def _record_id(version: str, index: int) -> str:
+    if version == "v2":
+        return f"curated-v2-synthetic-{index + 1:04d}"
+    return f"curated-synthetic-{index + 1:04d}"
+
+
+def _curation_notes(version: str) -> str:
+    if version == "v2":
+        return (
+            "Synthetic curated v2 row: no private photo, no personal identifier, "
+            "broader object and scene coverage, English-first output with Chinese helper text."
+        )
+    return (
+        "Synthetic curated row: no private photo, no personal identifier, "
+        "English-first output with Chinese helper text."
+    )
 
 
 def _build_object_understanding(obj: Mapping[str, object]) -> ObjectUnderstanding:
@@ -212,18 +482,19 @@ def _build_diary(obj: Mapping[str, object], mode: str, persona: Persona, index:
     profile = MODE_PROFILES[mode]
     object_name = str(obj["name"])
     features = ", ".join(str(feature) for feature in obj["features"][:2])
+    scene = str(obj.get("scene_detail", "collecting proof that ordinary things notice everything"))
     day_number = 300 + index + len(object_name)
     english = (
         f"Today I waited in the {obj['context']} wearing my {features} like official records. "
         f"The humans moved around me with the confidence of temporary weather. "
         f"I remembered how I {obj['memory']}, and I answered in my own way: {profile['voice']}. "
-        f"My mood is {persona.mood}, but I am still here, collecting proof that ordinary things notice everything."
+        f"My mood is {persona.mood}, but I am still here, {scene}."
     )
     chinese = (
         f"今天我待在 {obj['context']}，带着 {features}，像一份安静的档案。"
         f"人类从我身边经过，好像自己不是短暂天气。"
         f"我记得自己曾经 {obj['memory']}，于是用自己的方式回应：{profile['voice']}。"
-        f"我的情绪是 {persona.mood}，但我仍在这里，记录普通物品也会注意到的一切。"
+        f"我的情绪是 {persona.mood}，但我仍在这里，{scene}。"
     )
     return DiaryEntry(
         title=f"Secret Diary - Day {day_number}",
@@ -246,7 +517,10 @@ def _character_name(object_name: str, mode: str) -> str:
 
 def _object_description(obj: Mapping[str, object]) -> str:
     features = ", ".join(str(feature) for feature in obj["features"])
-    return f"{obj['name']} in a {obj['context']} with {features}"
+    description = f"{obj['name']} in a {obj['context']} with {features}"
+    if "scene_detail" in obj:
+        description = f"{description}, {obj['scene_detail']}"
+    return description
 
 
 def _user_prompt(object_understanding: Mapping[str, object], mode: str) -> str:
@@ -260,15 +534,17 @@ def _user_prompt(object_understanding: Mapping[str, object], mode: str) -> str:
 
 def _parse_args() -> argparse.Namespace:
     parser = argparse.ArgumentParser(description=__doc__)
-    parser.add_argument("--count", type=int, default=DEFAULT_COUNT)
-    parser.add_argument("--output", type=Path, default=DEFAULT_OUTPUT_PATH)
+    parser.add_argument("--version", choices=("v1", "v2"), default="v1")
+    parser.add_argument("--count", type=int, default=None)
+    parser.add_argument("--output", type=Path, default=None)
     return parser.parse_args()
 
 
 def main() -> None:
     args = _parse_args()
-    output_path = prepare_curated_dataset(args.output, args.count)
-    print(f"wrote {args.count} synthetic curated SFT records to {output_path}")
+    output_path = prepare_curated_dataset(args.output, args.count, version=args.version)
+    record_count = args.count or (DEFAULT_V2_COUNT if args.version == "v2" else DEFAULT_COUNT)
+    print(f"wrote {record_count} synthetic curated SFT records to {output_path}")
 
 
 if __name__ == "__main__":

scripts/publish_hf_dataset.py

@@ -0,0 +1,155 @@
+"""Upload a curated Objectverse Diary SFT JSONL file to Hugging Face Datasets."""
+
+from __future__ import annotations
+
+import argparse
+import json
+from pathlib import Path
+from typing import Any
+
+
+def validate_dataset_file(dataset_file: Path) -> dict[str, object]:
+    if not dataset_file.exists() or not dataset_file.is_file():
+        raise FileNotFoundError(f"Dataset file does not exist: {dataset_file}")
+
+    record_count = 0
+    sources: set[str] = set()
+    modes: set[str] = set()
+    object_names: set[str] = set()
+    for line_number, line in enumerate(
+        dataset_file.read_text(encoding="utf-8").splitlines(),
+        start=1,
+    ):
+        if not line.strip():
+            continue
+        try:
+            record = json.loads(line)
+        except json.JSONDecodeError as exc:
+            raise ValueError(f"Invalid JSON on line {line_number}: {exc.msg}") from exc
+        if not isinstance(record, dict):
+            raise ValueError(f"Line {line_number} must be a JSON object.")
+        messages = record.get("messages")
+        if not isinstance(messages, list) or not messages:
+            raise ValueError(f"Line {line_number} must include a non-empty messages list.")
+        assistant_messages = [
+            message
+            for message in messages
+            if isinstance(message, dict) and message.get("role") == "assistant"
+        ]
+        if not assistant_messages:
+            raise ValueError(f"Line {line_number} must include an assistant message.")
+        assistant_content = assistant_messages[-1].get("content")
+        if not isinstance(assistant_content, str):
+            raise ValueError(f"Line {line_number} assistant content must be a string.")
+        try:
+            assistant_payload = json.loads(assistant_content)
+        except json.JSONDecodeError as exc:
+            raise ValueError(
+                f"Line {line_number} assistant content is not valid JSON: {exc.msg}"
+            ) from exc
+        if not isinstance(assistant_payload, dict):
+            raise ValueError(f"Line {line_number} assistant content must be a JSON object.")
+        if "persona" not in assistant_payload or "diary" not in assistant_payload:
+            raise ValueError(
+                f"Line {line_number} assistant content must include persona and diary."
+            )
+
+        record_count += 1
+        if isinstance(record.get("source"), str):
+            sources.add(str(record["source"]))
+        if isinstance(record.get("mode"), str):
+            modes.add(str(record["mode"]))
+        object_understanding = record.get("object_understanding")
+        if isinstance(object_understanding, dict):
+            raw_object = object_understanding.get("object")
+            if isinstance(raw_object, dict) and isinstance(raw_object.get("name"), str):
+                object_names.add(str(raw_object["name"]))
+
+    if record_count == 0:
+        raise ValueError(f"Dataset file has no records: {dataset_file}")
+
+    return {
+        "dataset_file": str(dataset_file),
+        "record_count": record_count,
+        "sources": sorted(sources),
+        "modes": sorted(modes),
+        "unique_object_count": len(object_names),
+    }
+
+
+def upload_dataset(
+    *,
+    dataset_file: Path,
+    repo_id: str,
+    path_in_repo: str,
+    private: bool,
+    commit_message: str,
+    dry_run: bool,
+) -> dict[str, object]:
+    summary = validate_dataset_file(dataset_file)
+    summary.update(
+        {
+            "repo_id": repo_id,
+            "path_in_repo": path_in_repo,
+            "private": private,
+            "commit_message": commit_message,
+            "dry_run": dry_run,
+        }
+    )
+    if dry_run:
+        summary["uploaded"] = False
+        return summary
+
+    from huggingface_hub import HfApi
+
+    api = HfApi()
+    api.create_repo(repo_id=repo_id, repo_type="dataset", private=private, exist_ok=True)
+    api.upload_file(
+        path_or_fileobj=str(dataset_file),
+        path_in_repo=path_in_repo,
+        repo_id=repo_id,
+        repo_type="dataset",
+        commit_message=commit_message,
+    )
+    summary["uploaded"] = True
+    summary["url"] = f"https://huggingface.co/datasets/{repo_id}"
+    return summary
+
+
+def _print_json(payload: dict[str, Any]) -> None:
+    print(json.dumps(payload, indent=2, sort_keys=True), flush=True)
+
+
+def _parse_args() -> argparse.Namespace:
+    parser = argparse.ArgumentParser(description=__doc__)
+    parser.add_argument("--dataset-file", type=Path, required=True)
+    parser.add_argument("--repo-id", required=True)
+    parser.add_argument("--path-in-repo", required=True)
+    parser.add_argument("--private", action="store_true")
+    parser.add_argument(
+        "--commit-message",
+        default="Upload Objectverse Diary curated SFT dataset",
+    )
+    parser.add_argument("--dry-run", action="store_true")
+    return parser.parse_args()
+
+
+def main() -> None:
+    args = _parse_args()
+    _print_json(
+        upload_dataset(
+            dataset_file=args.dataset_file,
+            repo_id=args.repo_id,
+            path_in_repo=args.path_in_repo,
+            private=args.private,
+            commit_message=args.commit_message,
+            dry_run=args.dry_run,
+        )
+    )
+
+
+if __name__ == "__main__":
+    try:
+        main()
+    except Exception as exc:
+        raise SystemExit(str(exc)) from exc

scripts/publish_hf_gguf.py

@@ -0,0 +1,100 @@
+"""Upload an Objectverse Diary GGUF model file to Hugging Face Hub."""
+
+from __future__ import annotations
+
+import argparse
+import json
+from pathlib import Path
+from typing import Any
+
+
+def validate_gguf_file(gguf_file: Path) -> dict[str, object]:
+    if not gguf_file.exists() or not gguf_file.is_file():
+        raise FileNotFoundError(f"GGUF file does not exist: {gguf_file}")
+    if gguf_file.suffix.lower() != ".gguf":
+        raise ValueError(f"GGUF file must use .gguf suffix: {gguf_file}")
+    size_bytes = gguf_file.stat().st_size
+    if size_bytes <= 0:
+        raise ValueError(f"GGUF file is empty: {gguf_file}")
+    return {
+        "gguf_file": str(gguf_file),
+        "size_bytes": size_bytes,
+    }
+
+
+def upload_gguf(
+    *,
+    gguf_file: Path,
+    repo_id: str,
+    path_in_repo: str,
+    private: bool,
+    commit_message: str,
+    dry_run: bool,
+) -> dict[str, object]:
+    summary = validate_gguf_file(gguf_file)
+    summary.update(
+        {
+            "repo_id": repo_id,
+            "path_in_repo": path_in_repo,
+            "private": private,
+            "commit_message": commit_message,
+            "dry_run": dry_run,
+        }
+    )
+    if dry_run:
+        summary["uploaded"] = False
+        return summary
+
+    from huggingface_hub import HfApi
+
+    api = HfApi()
+    api.create_repo(repo_id=repo_id, repo_type="model", private=private, exist_ok=True)
+    api.upload_file(
+        path_or_fileobj=str(gguf_file),
+        path_in_repo=path_in_repo,
+        repo_id=repo_id,
+        repo_type="model",
+        commit_message=commit_message,
+    )
+    summary["uploaded"] = True
+    summary["url"] = f"https://huggingface.co/{repo_id}/blob/main/{path_in_repo}"
+    return summary
+
+
+def _print_json(payload: dict[str, Any]) -> None:
+    print(json.dumps(payload, indent=2, sort_keys=True), flush=True)
+
+
+def _parse_args() -> argparse.Namespace:
+    parser = argparse.ArgumentParser(description=__doc__)
+    parser.add_argument("--gguf-file", type=Path, required=True)
+    parser.add_argument("--repo-id", required=True)
+    parser.add_argument("--path-in-repo", required=True)
+    parser.add_argument("--private", action="store_true")
+    parser.add_argument(
+        "--commit-message",
+        default="Upload Objectverse Diary GGUF model",
+    )
+    parser.add_argument("--dry-run", action="store_true")
+    return parser.parse_args()
+
+
+def main() -> None:
+    args = _parse_args()
+    _print_json(
+        upload_gguf(
+            gguf_file=args.gguf_file,
+            repo_id=args.repo_id,
+            path_in_repo=args.path_in_repo,
+            private=args.private,
+            commit_message=args.commit_message,
+            dry_run=args.dry_run,
+        )
+    )
+
+
+if __name__ == "__main__":
+    try:
+        main()
+    except Exception as exc:
+        raise SystemExit(str(exc)) from exc

src/examples.py

@@ -52,8 +52,9 @@ def gradio_examples() -> list[list[str]]:
 
 def example_button_label(index: int) -> str:
     item = EXAMPLE_OBJECTS[index]
+    label = item["label"].split("/")[0].strip()
     return (
         f"{item['archive_id']}\n"
-        f"{item['label']}\n"
+        f"{label}\n"
         f"{item['mode']} · {item['tags']}"
     )

src/renderer/share_card.py

@@ -15,14 +15,14 @@ def render_share_card(persona: PersonaEnvelope, diary: DiaryEntry) -> str:
     <article class="{CARD_WRAPPER_CLASS}">
       <header class="card-header">
         <div>
-          <div class="card-kicker">Objectverse Diary / 万物日记</div>
+          <div class="card-kicker">Objectverse Diary <span class="lang-zh">万物日记</span></div>
           <h2>{escape(p.character_name)}</h2>
         </div>
         <span class="card-stamp">OBJECT FILE</span>
       </header>
       <p class="card-object">{escape(p.object_name)} · {escape(p.mood)}</p>
       <p class="card-quote">{escape(diary.english)}</p>
-      <p class="card-cn">{escape(diary.chinese)}</p>
+      <p class="card-cn lang-zh block">{escape(diary.chinese)}</p>
       <div class="card-tags">{tags}</div>
     </article>
     """

src/ui/copy.py

@@ -1,19 +1,19 @@
-"""English-first, Chinese-second UI copy."""
+"""English UI copy."""
 
 TITLE = "Objectverse Diary"
-SUBTITLE = "Every object has a secret life. / 每个物品都有秘密人生。"
-UPLOAD_LABEL = "Upload an object photo / 上传一个物品照片"
-DESCRIPTION_LABEL = "Optional object description / 可选物品描述"
+SUBTITLE = "Every object has a secret life."
+UPLOAD_LABEL = "Upload an object photo"
+DESCRIPTION_LABEL = "Optional object description"
 DESCRIPTION_PLACEHOLDER = "Example: an old white coffee mug on my developer desk"
-MODE_LABEL = "Choose a personality mode / 选择人格模式"
-GENERATE_LABEL = "Wake the object / 唤醒这个物品"
-EXAMPLES_LABEL = "Example objects / 示例物品"
-OBJECT_JSON_LABEL = "Object Understanding JSON / 物品识别 JSON"
-PERSONA_JSON_LABEL = "Persona JSON / 人格 JSON"
-DIARY_LABEL = "Secret Diary / 秘密日记"
-SHARE_CARD_LABEL = "Share Card / 分享卡片"
-TRACE_JSON_LABEL = "Trace JSON / 模型轨迹 JSON"
-TRACE_PATH_LABEL = "Saved trace path / 已保存 trace 路径"
-CHAT_LABEL = "Chat with the object / 和物品对话"
+MODE_LABEL = "Choose a personality mode"
+GENERATE_LABEL = "Wake the object"
+EXAMPLES_LABEL = "Example objects"
+OBJECT_JSON_LABEL = "Object Understanding JSON"
+PERSONA_JSON_LABEL = "Persona JSON"
+DIARY_LABEL = "Secret Diary"
+SHARE_CARD_LABEL = "Share Card"
+TRACE_JSON_LABEL = "Trace JSON"
+TRACE_PATH_LABEL = "Saved trace path"
+CHAT_LABEL = "Chat with the object"
 CHAT_INPUT_PLACEHOLDER = "Ask the object something..."
-CHAT_BUTTON_LABEL = "Ask / 追问"
+CHAT_BUTTON_LABEL = "Ask"

src/ui/layout.py

@@ -19,37 +19,194 @@ from src.renderer.share_card import render_share_card
 from src.ui import copy
 from src.utils.zero_gpu import zero_gpu
 
-CHAT_EMPTY_MESSAGE = "Wake an object first. / 请先唤醒一个物品。"
+CHAT_EMPTY_MESSAGE = "Wake an object first."
 
 OBJECT_FILE_EMPTY = """
 <div class="archive-empty">
-  <span class="archive-label">Object File / 物品档案</span>
+  <span class="archive-label">Object File <span class="lang-zh">物品档案</span></span>
   <h3>No object awake yet.</h3>
-  <p>Upload or describe an everyday object to open its secret archive. / 上传或描述一个日常物品后打开秘密档案。</p>
+  <p>Upload or describe an everyday object to open its secret archive.</p>
+  <p class="lang-zh block">上传或描述一个日常物品后打开秘密档案。</p>
 </div>
 """
 
 DIARY_EMPTY = """
-### Secret Diary / 秘密日记
+### Secret Diary
 
-Wake an object to open its diary. / 唤醒物品后阅读它的日记。
+Wake an object to open its diary.
+
+<div class="lang-zh block zh-helper">
+唤醒物品后阅读它的日记。
+</div>
 """
 
 SHARE_CARD_EMPTY = """
 <div class="objectverse-placeholder">
-  <span>Share Card / 分享卡片</span>
+  <span>Share Card <span class="lang-zh">分享卡片</span></span>
   <strong>Waiting for an object file.</strong>
-  <p>A screenshot-friendly archive card will appear here. / 可截图分享的档案卡片会显示在这里。</p>
+  <p>A screenshot-friendly archive card will appear here.</p>
+  <p class="lang-zh block">可截图分享的档案卡片会显示在这里。</p>
 </div>
 """
 
 TRACE_EMPTY = """
 <div class="archive-empty compact">
-  <span class="archive-label">Trace / 模型轨迹</span>
-  <p>No trace saved yet. / 尚未保存 trace。</p>
+  <span class="archive-label">Trace <span class="lang-zh">模型轨迹</span></span>
+  <p>No trace saved yet.</p>
+  <p class="lang-zh block">尚未保存 trace。</p>
 </div>
 """
 
+UI_CONTROL_SCRIPT = r"""
+(() => {
+  const root = document.documentElement;
+  const INTERNAL_TEXT_REPLACEMENTS = new Map([
+    ["将图像文件拖放到此处以上传", "Drop image file here to upload"],
+    ["将图像拖放到此处", "Drop image here"],
+    ["- 或 -", "- or -"],
+    ["点击上传", "Click to upload"],
+    ["清空对话", "Clear chat"],
+    ["通过 API 使用", "Use via API"],
+    ["使用 Gradio 构建", "Built with Gradio"],
+    ["设置", "Settings"],
+    ["标志", "icon"],
+  ]);
+  const CJK_RE = /[\u3400-\u9fff]/;
+  const CJK_WRAP_RE = /[\u3400-\u9fff，。！？、；：：“”‘’（）《》【】]+/g;
+  const SKIP_TEXT_SELECTOR = "script, style, textarea, input, select, option, svg, .lang-zh, .auto-zh";
+
+  function syncLanguageButtons(value) {
+    document.querySelectorAll("[data-lang-toggle]").forEach((button) => {
+      const active = button.dataset.langToggle === value;
+      button.classList.toggle("active", active);
+      button.setAttribute("aria-pressed", String(active));
+    });
+  }
+
+  function applyLanguage(value) {
+    const language = value === "zh" ? "zh" : "en";
+    root.dataset.ovLang = language;
+    if (document.body) {
+      document.body.dataset.ovLang = language;
+    }
+    syncLanguageButtons(language);
+  }
+
+  function initControls() {
+    root.lang = "en";
+    applyLanguage("en");
+    normalizeGradioChrome(document.body);
+    wrapChineseText(document.body);
+  }
+
+  function normalizeString(value) {
+    let nextValue = value;
+    INTERNAL_TEXT_REPLACEMENTS.forEach((replacement, source) => {
+      nextValue = nextValue.split(source).join(replacement);
+    });
+    return nextValue;
+  }
+
+  function normalizeGradioChrome(rootNode) {
+    if (!rootNode) return;
+
+    rootNode.querySelectorAll("[aria-label], [title], [alt]").forEach((element) => {
+      ["aria-label", "title", "alt"].forEach((attribute) => {
+        const value = element.getAttribute(attribute);
+        if (value && CJK_RE.test(value)) {
+          const normalizedValue = normalizeString(value);
+          if (normalizedValue !== value) {
+            element.setAttribute(attribute, normalizedValue);
+          }
+        }
+      });
+    });
+
+    const walker = document.createTreeWalker(rootNode, NodeFilter.SHOW_TEXT);
+    const nodes = [];
+    let node = walker.nextNode();
+    while (node) {
+      const parent = node.parentElement;
+      const text = node.nodeValue || "";
+      if (parent && !parent.closest(SKIP_TEXT_SELECTOR) && CJK_RE.test(text)) {
+        nodes.push(node);
+      }
+      node = walker.nextNode();
+    }
+
+    nodes.forEach((textNode) => {
+      const text = textNode.nodeValue || "";
+      const normalizedText = normalizeString(text);
+      if (normalizedText !== text) {
+        textNode.nodeValue = normalizedText;
+      }
+    });
+  }
+
+  function wrapChineseText(rootNode) {
+    if (!rootNode) return;
+
+    const walker = document.createTreeWalker(rootNode, NodeFilter.SHOW_TEXT);
+    const nodes = [];
+    let node = walker.nextNode();
+    while (node) {
+      const parent = node.parentElement;
+      const text = node.nodeValue || "";
+      if (parent && !parent.closest(SKIP_TEXT_SELECTOR) && CJK_RE.test(text)) {
+        nodes.push(node);
+      }
+      node = walker.nextNode();
+    }
+
+    nodes.forEach((textNode) => {
+      const text = textNode.nodeValue || "";
+      const fragment = document.createDocumentFragment();
+      let lastIndex = 0;
+      text.replace(CJK_WRAP_RE, (match, index) => {
+        if (index > lastIndex) {
+          fragment.append(document.createTextNode(text.slice(lastIndex, index)));
+        }
+        const span = document.createElement("span");
+        span.className = "auto-zh";
+        span.textContent = match;
+        fragment.append(span);
+        lastIndex = index + match.length;
+        return match;
+      });
+      if (lastIndex < text.length) {
+        fragment.append(document.createTextNode(text.slice(lastIndex)));
+      }
+      textNode.replaceWith(fragment);
+    });
+  }
+
+  document.addEventListener("click", (event) => {
+    const langButton = event.target.closest("[data-lang-toggle]");
+    if (langButton) {
+      applyLanguage(langButton.dataset.langToggle);
+    }
+  });
+
+  if (document.readyState === "loading") {
+    document.addEventListener("DOMContentLoaded", initControls);
+  } else {
+    initControls();
+  }
+
+  const observer = new MutationObserver(() => {
+    normalizeGradioChrome(document.body);
+    wrapChineseText(document.body);
+  });
+  if (document.body) {
+    observer.observe(document.body, { childList: true, subtree: true });
+  } else {
+    document.addEventListener("DOMContentLoaded", () => {
+      observer.observe(document.body, { childList: true, subtree: true });
+    });
+  }
+})();
+"""
+
 GenerationUiResult = tuple[
     str,
     dict[str, Any],
@@ -80,210 +237,194 @@ def build_app() -> gr.Blocks:
         background_fill_secondary_dark="rgba(30, 28, 25, 0.6)",
         border_color_primary="rgba(212, 175, 55, 0.15)",
         border_color_primary_dark="rgba(212, 175, 55, 0.15)",
+        body_text_color="#E6E1D3",
+        body_text_color_dark="#E6E1D3",
+        body_text_color_subdued="#A89B84",
+        body_text_color_subdued_dark="#A89B84",
+        link_text_color="#D4AF37",
+        link_text_color_dark="#D4AF37",
+        link_text_color_hover="#F5D061",
+        link_text_color_hover_dark="#F5D061",
+        link_text_color_active="#F5D061",
+        link_text_color_active_dark="#F5D061",
+        link_text_color_visited="#D4AF37",
+        link_text_color_visited_dark="#D4AF37",
         block_background_fill="transparent",
         block_background_fill_dark="transparent",
         block_border_width="0px",
+        block_info_text_color="#A89B84",
+        block_info_text_color_dark="#A89B84",
+        block_label_text_color="#A89B84",
+        block_label_text_color_dark="#A89B84",
+        block_title_text_color="#E6E1D3",
+        block_title_text_color_dark="#E6E1D3",
         panel_background_fill="transparent",
         panel_background_fill_dark="transparent",
+        accordion_text_color="#E6E1D3",
+        accordion_text_color_dark="#E6E1D3",
+        table_text_color="#E6E1D3",
+        table_text_color_dark="#E6E1D3",
+        input_background_fill="#1b1a18",
+        input_background_fill_dark="#1b1a18",
+        input_background_fill_focus="#1b1a18",
+        input_background_fill_focus_dark="#1b1a18",
+        input_background_fill_hover="#1b1a18",
+        input_background_fill_hover_dark="#1b1a18",
+        input_border_color="rgba(212, 175, 55, 0.3)",
+        input_border_color_dark="rgba(212, 175, 55, 0.3)",
+        input_border_color_focus="#D4AF37",
+        input_border_color_focus_dark="#D4AF37",
+        input_placeholder_color="#8B8678",
+        input_placeholder_color_dark="#8B8678",
+        checkbox_label_text_color="#E6E1D3",
+        checkbox_label_text_color_dark="#E6E1D3",
+        checkbox_label_text_color_selected="#F5D061",
+        checkbox_label_text_color_selected_dark="#F5D061",
+        checkbox_label_background_fill="transparent",
+        checkbox_label_background_fill_dark="transparent",
+        checkbox_label_background_fill_selected="rgba(212, 175, 55, 0.05)",
+        checkbox_label_background_fill_selected_dark="rgba(212, 175, 55, 0.05)",
+        checkbox_label_border_color="rgba(212, 175, 55, 0.3)",
+        checkbox_label_border_color_dark="rgba(212, 175, 55, 0.3)",
+        checkbox_label_border_color_selected="#D4AF37",
+        checkbox_label_border_color_selected_dark="#D4AF37",
+        button_secondary_background_fill="rgba(22, 21, 19, 0.8)",
+        button_secondary_background_fill_dark="rgba(22, 21, 19, 0.8)",
+        button_secondary_background_fill_hover="rgba(38, 35, 29, 0.9)",
+        button_secondary_background_fill_hover_dark="rgba(38, 35, 29, 0.9)",
+        button_secondary_border_color="rgba(212, 175, 55, 0.15)",
+        button_secondary_border_color_dark="rgba(212, 175, 55, 0.15)",
+        button_secondary_border_color_hover="#D4AF37",
+        button_secondary_border_color_hover_dark="#D4AF37",
+        button_secondary_text_color="#E6E1D3",
+        button_secondary_text_color_dark="#E6E1D3",
+        button_secondary_text_color_hover="#F5D061",
+        button_secondary_text_color_hover_dark="#F5D061",
+        button_primary_text_color="#2a261f",
+        button_primary_text_color_dark="#2a261f",
     )
 
-    with gr.Blocks(theme=custom_theme, head=f"<style>{css}</style>", title=APP_TITLE, fill_width=True, elem_id="objectverse-app") as demo:
-        with gr.Row(elem_id="app-container"):
-            # === Sidebar ===
-            with gr.Column(elem_id="sidebar", scale=0, min_width=240):
-                gr.HTML(
-                    """
-                    <nav class="sidebar-nav">
-                      <div class="sidebar-logo">
-                        <div class="logo-icon"></div>
-                        <h2>Objectverse<br>Diary</h2>
-                      </div>
-                      <ul class="sidebar-menu">
-                        <li class="active"><a href="#intake">Home</a></li>
-                        <li><a href="#intake">Intake</a></li>
-                        <li><a href="#object-file">Object File</a></li>
-                        <li><a href="#diary">Diary</a></li>
-                        <li><a href="#chat-panel">Chat</a></li>
-                        <li><a href="#share-panel">Share Card</a></li>
-                        <li><a href="#trace">Trace</a></li>
-                        <li><a href="#settings">Settings</a></li>
-                      </ul>
-                      <div class="sidebar-footer">
-                        <div class="footer-stamp">
-                          <small>OBJECTVERSE ARCHIVE</small>
-                          <span>No. 000827</span>
-                          <small>Curate. Converse. Cherish.</small>
-                        </div>
-                        <div class="lang-switch">
-                          <button class="active">EN</button>
-                          <button>中文</button>
-                        </div>
-                      </div>
-                    </nav>
-                    """,
-                    padding=False,
-                )
-
-            # === Main Content Area ===
-            with gr.Column(elem_id="main-content", scale=1):
-                gr.HTML(
-                    f"""
-                    <section id="objectverse-hero">
-                        <div class="hero-copy">
-                          <h1>{APP_TITLE}</h1>
-                          <p class="hero-kicker">Every object has a secret life.<br><span>万物日记：每个物品都有秘密人生</span></p>
-                        </div>
-                        <div class="hero-badges" aria-label="Project constraints">
-                          <span>Small Models</span>
-                          <span>Local-First</span>
-                          <span>No Cloud APIs</span>
-                        </div>
-                    </section>
-                    """,
-                    padding=False,
-                )
-
-                result_state = gr.State()
-                zero_gpu_probe_button = gr.Button(visible=False)
-                zero_gpu_probe_output = gr.JSON(visible=False)
-                vision_runtime_probe_button = gr.Button(visible=False)
-                vision_runtime_probe_output = gr.JSON(visible=False)
-
-                # Intake & Examples Row
-                with gr.Row(elem_id="intake", elem_classes=["content-section"]):
-                    # Left: Intake
-                    with gr.Column(scale=7, elem_classes=["archive-panel", "intake-panel"]):
-                        image_input = gr.Image(
-                            label=copy.UPLOAD_LABEL,
-                            show_label=False,
-                            type="filepath",
-                            sources=["upload"],
-                            elem_id="object-upload",
-                        )
-                        gr.HTML("""<div class="or-divider"><span>OR</span></div>""", padding=False)
-                        description_input = gr.Textbox(
-                            label=copy.DESCRIPTION_LABEL,
-                            placeholder=copy.DESCRIPTION_PLACEHOLDER,
-                            lines=2,
-                            max_lines=5,
-                            elem_id="object-description",
-                        )
-                        
-                        gr.HTML("""<div class="mode-header">Personality mode <small>人格模式</small> <span class="help-icon">?</span></div>""", padding=False)
-                        mode_input = gr.Radio(
-                            label=copy.MODE_LABEL,
-                            show_label=False,
-                            choices=PERSONALITY_MODES,
-                            value=DEFAULT_MODE,
-                            elem_id="personality-mode",
-                            elem_classes=["mode-switch"],
-                        )
-                        generate_button = gr.Button("Wake the Object\n唤醒物品", variant="primary", elem_id="wake-button")
-
-                        gr.HTML(
-                            """
-                            <div class="how-it-works">
-                              <div class="step">
-                                <span class="step-num">01</span>
-                                <div class="step-icon img-icon"></div>
-                                <div class="step-text">
-                                  <strong>Upload or describe</strong>
-                                  <small>上传物品或描述心情</small>
-                                  <p>Give me a photo or words—anything that holds a story.</p>
-                                </div>
-                              </div>
-                              <div class="step">
-                                <span class="step-num">02</span>
-                                <div class="step-icon pen-icon"></div>
-                                <div class="step-text">
-                                  <strong>I imagine its life</strong>
-                                  <small>我为它编织人生</small>
-                                  <p>I'll step into its shoes and imagine its secret life.</p>
-                                </div>
-                              </div>
-                              <div class="step">
-                                <span class="step-num">03</span>
-                                <div class="step-icon book-icon"></div>
-                                <div class="step-text">
-                                  <strong>Read its diary</strong>
-                                  <small>阅读物品日记</small>
-                                  <p>Receive a diary entry written from its perspective.</p>
-                                </div>
-                              </div>
-                            </div>
-                            """,
-                            padding=False,
-                        )
+    with gr.Blocks(theme=custom_theme, head=f"<style>{css}</style><script>{UI_CONTROL_SCRIPT}</script>", title=APP_TITLE, fill_width=True, elem_id="objectverse-app") as demo:
+        with gr.Column(elem_id="app-container"):
+            gr.HTML(
+                f"""
+                <header id="objectverse-hero">
+                  <div class="hero-copy">
+                    <span class="archive-label">Small Model Object Archive</span>
+                    <h1>{APP_TITLE}</h1>
+                    <p class="hero-kicker">Every object has a secret life.</p>
+                    <p class="hero-feature">Build Small Hackathon entry: upload an object, wake its secret persona, read the diary, chat, and share the evidence. Tiny models, weird lives.</p>
+                    <p class="hero-kicker lang-zh block">万物日记：每个物品都有秘密人生。</p>
+                    <p class="hero-feature lang-zh block">Build Small 黑客松作品：上传物品，唤醒隐藏人格，读日记、追问它，再分享证据。小模型，怪人生。</p>
+                  </div>
+                  <div class="top-controls" aria-label="Display controls">
+                    <span>Language</span>
+                    <div class="segmented-control">
+                      <button type="button" class="active" data-lang-toggle="en" aria-pressed="true">EN</button>
+                      <button type="button" data-lang-toggle="zh" aria-pressed="false">ZH</button>
+                    </div>
+                  </div>
+                </header>
+                """,
+                padding=False,
+            )
 
-                    # Right: Examples
-                    with gr.Column(scale=4, elem_classes=["archive-panel", "examples-panel"]):
-                        gr.HTML(
-                            """
-                            <div class="example-header">
-                              <div class="books-icon"></div>
-                              <div>
-                                <strong>Example Objects</strong>
-                                <span>灵感库</span>
-                              </div>
-                            </div>
-                            """,
-                            padding=False,
-                        )
-                        example_buttons: list[gr.Button] = []
-                        for index in range(len(EXAMPLE_OBJECTS)):
-                            example_buttons.append(
-                                gr.Button(
-                                    example_button_label(index),
-                                    elem_classes=["example-card"],
-                                    variant="secondary",
-                                )
+            result_state = gr.State()
+            zero_gpu_probe_button = gr.Button(visible=False)
+            zero_gpu_probe_output = gr.JSON(visible=False)
+            vision_runtime_probe_button = gr.Button(visible=False)
+            vision_runtime_probe_output = gr.JSON(visible=False)
+
+            with gr.Row(elem_id="intake", elem_classes=["content-section", "top-grid"]):
+                with gr.Column(scale=7, elem_classes=["archive-panel", "intake-panel"]):
+                    gr.HTML(_panel_header("01", "Wake an Object", "Upload a photo or describe an everyday object.", "唤醒物品"), padding=False)
+                    image_input = gr.Image(
+                        label=copy.UPLOAD_LABEL,
+                        show_label=False,
+                        type="filepath",
+                        sources=["upload"],
+                        placeholder="Drop an object photo here or click to upload.",
+                        elem_id="object-upload",
+                    )
+                    gr.HTML("""<div class="or-divider"><span>OR</span></div>""", padding=False)
+                    description_input = gr.Textbox(
+                        label=copy.DESCRIPTION_LABEL,
+                        placeholder=copy.DESCRIPTION_PLACEHOLDER,
+                        lines=2,
+                        max_lines=5,
+                        elem_id="object-description",
+                    )
+
+                    gr.HTML("""<div class="mode-header">Personality mode <span class="lang-zh">人格模式</span></div>""", padding=False)
+                    mode_input = gr.Radio(
+                        label=copy.MODE_LABEL,
+                        show_label=False,
+                        choices=PERSONALITY_MODES,
+                        value=DEFAULT_MODE,
+                        elem_id="personality-mode",
+                        elem_classes=["mode-switch"],
+                    )
+                    generate_button = gr.Button("Wake the Object", variant="primary", elem_id="wake-button")
+
+                with gr.Column(scale=4, elem_classes=["archive-panel", "examples-panel"]):
+                    gr.HTML(
+                        """
+                        <div class="example-header">
+                          <div>
+                            <strong>Example Objects</strong>
+                            <span class="lang-zh block">示例物品</span>
+                          </div>
+                        </div>
+                        """,
+                        padding=False,
+                    )
+                    example_buttons: list[gr.Button] = []
+                    for index in range(len(EXAMPLE_OBJECTS)):
+                        example_buttons.append(
+                            gr.Button(
+                                example_button_label(index),
+                                elem_classes=["example-card"],
+                                variant="secondary",
                             )
-                        gr.HTML("""<a href="#object-file" class="view-more">View more in Object File →</a>""", padding=False)
-
-                # Object File Section
-                with gr.Row(elem_id="object-file", elem_classes=["content-section"]):
-                    with gr.Column(scale=1, elem_classes=["archive-panel", "file-panel"]):
-                        gr.HTML(_panel_header("02", "Object File / Recognition", "物品档案", "Structured mock understanding and persona."), padding=False)
-                        object_file_summary = gr.HTML(value=OBJECT_FILE_EMPTY, elem_id="object-file-summary", padding=False)
-                        with gr.Accordion("Raw JSON", open=False):
-                            object_json = gr.JSON(value={}, label=copy.OBJECT_JSON_LABEL)
-                            persona_json = gr.JSON(value={}, label=copy.PERSONA_JSON_LABEL)
-
-                # Diary Section
-                with gr.Row(elem_id="diary", elem_classes=["content-section"]):
-                    with gr.Column(scale=1, elem_classes=["archive-panel", "diary-panel"]):
-                        gr.HTML(_panel_header("03", "Secret Diary", "秘密日记", "A private note written by the object."), padding=False)
-                        diary_output = gr.Markdown(
-                            value=DIARY_EMPTY,
-                            label=copy.DIARY_LABEL,
-                            elem_id="diary-output",
-                        )
-
-                # Share & Chat Section
-                with gr.Row(elem_id="share", elem_classes=["content-section", "split-section"]):
-                    with gr.Column(scale=5, elem_classes=["archive-panel", "share-panel", "anchored"], elem_id="share-panel"):
-                        gr.HTML(_panel_header("04", "Share Card", "分享卡片", "Fixed-width card for screenshots."), padding=False)
-                        share_card = gr.HTML(value=SHARE_CARD_EMPTY, label=copy.SHARE_CARD_LABEL, padding=False)
-
-                    with gr.Column(scale=4, elem_classes=["archive-panel", "chat-panel", "anchored"], elem_id="chat-panel"):
-                        gr.HTML(_panel_header("05", "Object Chat", "物品对话", "Ask after the object wakes up."), padding=False)
-                        chatbot = gr.Chatbot(
-                            value=_empty_chat_history(),
-                            label=copy.CHAT_LABEL,
-                            type="messages",
-                            height=300,
-                            allow_tags=False,
                         )
-                        chat_input = gr.Textbox(placeholder=copy.CHAT_INPUT_PLACEHOLDER, show_label=False)
-                        chat_button = gr.Button(copy.CHAT_BUTTON_LABEL, elem_classes=["quiet-button"])
 
-                # Trace Section
-                with gr.Row(elem_id="trace", elem_classes=["content-section"]):
-                    with gr.Column(scale=1, elem_classes=["archive-panel", "trace-panel"]):
-                        gr.HTML(_panel_header("06", "Trace", "模型轨迹", "Saved JSON record for reproducibility."), padding=False)
-                        trace_summary = gr.HTML(value=TRACE_EMPTY, elem_id="trace-summary", padding=False)
-                        trace_json = gr.JSON(value={}, label=copy.TRACE_JSON_LABEL)
-                        trace_path = gr.Textbox(label=copy.TRACE_PATH_LABEL, interactive=False)
+            with gr.Row(elem_id="results", elem_classes=["content-section", "results-grid"]):
+                with gr.Column(scale=5, elem_classes=["archive-panel", "file-panel"]):
+                    gr.HTML(_panel_header("02", "Object File", "Structured understanding and persona.", "物品档案"), padding=False)
+                    object_file_summary = gr.HTML(value=OBJECT_FILE_EMPTY, elem_id="object-file-summary", padding=False)
+
+                with gr.Column(scale=6, elem_classes=["archive-panel", "diary-panel"]):
+                    gr.HTML(_panel_header("03", "Secret Diary", "A private note written by the object.", "秘密日记"), padding=False)
+                    diary_output = gr.Markdown(
+                        value=DIARY_EMPTY,
+                        label=copy.DIARY_LABEL,
+                        elem_id="diary-output",
+                    )
+
+            with gr.Row(elem_id="share-chat", elem_classes=["content-section", "split-section"]):
+                with gr.Column(scale=5, elem_classes=["archive-panel", "share-panel"], elem_id="share-panel"):
+                    gr.HTML(_panel_header("04", "Share Card", "Fixed-width card for screenshots.", "分享卡片"), padding=False)
+                    share_card = gr.HTML(value=SHARE_CARD_EMPTY, label=copy.SHARE_CARD_LABEL, padding=False)
+
+                with gr.Column(scale=4, elem_classes=["archive-panel", "chat-panel"], elem_id="chat-panel"):
+                    gr.HTML(_panel_header("05", "Object Chat", "Ask after the object wakes up.", "物品对话"), padding=False)
+                    chatbot = gr.Chatbot(
+                        value=_empty_chat_history(),
+                        label=copy.CHAT_LABEL,
+                        type="messages",
+                        height=300,
+                        allow_tags=False,
+                    )
+                    chat_input = gr.Textbox(placeholder=copy.CHAT_INPUT_PLACEHOLDER, show_label=False)
+                    chat_button = gr.Button(copy.CHAT_BUTTON_LABEL, elem_classes=["quiet-button"])
+
+            with gr.Accordion("Developer details", open=False, elem_classes=["developer-details"]):
+                trace_summary = gr.HTML(value=TRACE_EMPTY, elem_id="trace-summary", padding=False)
+                with gr.Row(elem_classes=["developer-json-grid"]):
+                    object_json = gr.JSON(value={}, label=copy.OBJECT_JSON_LABEL)
+                    persona_json = gr.JSON(value={}, label=copy.PERSONA_JSON_LABEL)
+                trace_json = gr.JSON(value={}, label=copy.TRACE_JSON_LABEL)
+                trace_path = gr.Textbox(label=copy.TRACE_PATH_LABEL, interactive=False)
 
         manual_outputs = [
             object_file_summary,
@@ -337,12 +478,13 @@ def build_app() -> gr.Blocks:
     return demo
 
 
-def _panel_header(index: str, title: str, chinese: str, note: str) -> str:
+def _panel_header(index: str, title: str, note: str, chinese: str = "") -> str:
+    chinese_label = f' <small class="lang-zh">{escape(chinese)}</small>' if chinese else ""
     return f"""
     <header class="panel-header">
       <span>{escape(index)}</span>
       <div>
-        <h2>{escape(title)} <small>{escape(chinese)}</small></h2>
+        <h2>{escape(title)}{chinese_label}</h2>
         <p>{escape(note)}</p>
       </div>
     </header>
@@ -422,7 +564,7 @@ def _render_object_file(result: GenerationResult) -> str:
         </div>
       </dl>
       <div class="feature-list">
-        <strong>Visible features / 可见特征</strong>
+        <strong>Visible features <span class="lang-zh">可见特征</span></strong>
         <ul>{features}</ul>
       </div>
       <p class="complaint">{escape(persona.complaint)}</p>
@@ -434,7 +576,7 @@ def _render_object_file(result: GenerationResult) -> str:
 def _render_trace_summary(result: GenerationResult) -> str:
     return f"""
     <div class="trace-card">
-      <span class="archive-label">Trace saved / Trace 已保存</span>
+      <span class="archive-label">Trace saved <span class="lang-zh">Trace 已保存</span></span>
       <strong>{escape(result.trace.trace_id)}</strong>
       <p>{escape(result.trace.model_runtime["vision"])} · {escape(result.trace.model_runtime["text"])}</p>
     </div>
@@ -451,15 +593,16 @@ def _generation_error(exc: Exception, description: str, mode: str) -> Generation
     }
     error_html = f"""
     <div class="archive-error">
-      <span>Generation failed / 生成失败</span>
+      <span>Generation failed <span class="lang-zh">生成失败</span></span>
       <strong>{escape(error_type)}</strong>
       <p>{escape(error_message)}</p>
     </div>
     """
     error_markdown = (
-        "### Generation failed / 生成失败\n\n"
+        "### Generation failed\n\n"
         f"{error_type}: {error_message}\n\n"
-        "Please try another description or sample object. / 请尝试其他描述或示例物品。"
+        "Please try another description or sample object.\n\n"
+        '<div class="lang-zh block zh-helper">请尝试其他描述或示例物品。</div>'
     )
     return (
         error_html,
@@ -471,7 +614,7 @@ def _generation_error(exc: Exception, description: str, mode: str) -> Generation
         error_payload,
         "",
         None,
-        [{"role": "assistant", "content": f"Generation failed. / 生成失败：{error_type}"}],
+        [{"role": "assistant", "content": f"Generation failed: {error_type}"}],
     )
 
 
@@ -484,7 +627,7 @@ def _awake_chat_history(result: GenerationResult) -> list[dict[str, str]]:
     return [
         {
             "role": "assistant",
-            "content": f"{name} is awake. Ask what it remembers. / {name} 已被唤醒，可以追问它记得什么。",
+            "content": f"{name} is awake. Ask what it remembers.",
         }
     ]
 

src/ui/styles.css

@@ -1,618 +1,649 @@
-/* 
- * Objectverse Diary - Dark Academia / Vintage Archive Theme 
- * Updated to match reference UI.
+/*
+ * Objectverse Diary - compact archive UI.
  */
 
- @import url('https://fonts.googleapis.com/css2?family=Space+Mono:ital,wght@0,400;0,700;1,400&family=Courier+Prime:ital,wght@0,400;0,700;1,400&display=swap');
-
- :root {
-   --ov-bg: #161513;
-   --ov-bg-panel: rgba(30, 28, 25, 0.6);
-   --ov-bg-input: #1b1a18;
-   
-   --ov-border-faint: rgba(212, 175, 55, 0.15);
-   --ov-border-light: rgba(212, 175, 55, 0.3);
-   --ov-border-strong: rgba(212, 175, 55, 0.8);
-   
-   --ov-text-main: #E6E1D3;
-   --ov-text-muted: #8B8678;
-   --ov-text-dark: #2a261f;
-   
-   --ov-gold: #D4AF37;
-   --ov-gold-bright: #F5D061;
-   
-   --font-typewriter: 'Courier Prime', 'Space Mono', 'Courier New', monospace;
-   --font-sans: 'Inter', -apple-system, sans-serif;
-   --font-serif: Georgia, serif;
- }
- 
- html, body, gradio-app {
-   background-color: var(--ov-bg);
-   margin: 0;
-   padding: 0;
-   width: 100%;
-   height: 100%;
-   color: var(--ov-text-main);
- }
- 
- /* Subtle noise overlay */
- body::before {
-   content: "";
-   position: fixed;
-   top: 0; left: 0; right: 0; bottom: 0;
-   background-image: url("data:image/svg+xml,%3Csvg viewBox='0 0 200 200' xmlns='http://www.w3.org/2000/svg'%3E%3Cfilter id='noiseFilter'%3E%3CfeTurbulence type='fractalNoise' baseFrequency='0.85' numOctaves='3' stitchTiles='stitch'/%3E%3C/filter%3E%3Crect width='100%25' height='100%25' filter='url(%23noiseFilter)' opacity='0.03'/%3E%3C/svg%3E");
-   pointer-events: none;
-   z-index: 9999;
- }
- 
- .gradio-container {
-   max-width: 100% !important;
-   padding: 0 !important;
-   background: transparent !important;
-   font-family: var(--font-sans);
- }
- 
- /* Layout wrapper */
- #app-container {
-   display: flex;
-   flex-direction: row;
-   min-height: 100vh;
-   align-items: stretch;
-   gap: 0 !important;
-   margin: 0 !important;
- }
- 
- /* ====================
-    Sidebar Styles
-    ==================== */
- #sidebar {
-   width: 240px;
-   min-width: 240px !important;
-   max-width: 240px !important;
-   border-right: 1px solid var(--ov-border-faint);
-   background: rgba(22, 21, 19, 0.95);
-   position: fixed;
-   top: 0;
-   bottom: 0;
-   left: 0;
-   display: flex;
-   flex-direction: column;
-   z-index: 100;
-   padding: 30px 0;
- }
- 
- .sidebar-logo {
-   text-align: center;
-   margin-bottom: 40px;
- }
- 
- .sidebar-logo h2 {
-   font-family: var(--font-typewriter);
-   font-size: 18px;
-   color: var(--ov-text-main);
-   margin: 10px 0 0;
-   line-height: 1.2;
-   font-weight: normal;
- }
- 
- .logo-icon {
-   width: 48px;
-   height: 64px;
-   margin: 0 auto;
-   border: 1px solid var(--ov-gold);
-   border-radius: 24px;
-   display: flex;
-   align-items: center;
-   justify-content: center;
-   position: relative;
- }
- 
- .logo-icon::after {
-   content: "⚷"; /* Key symbol placeholder */
-   color: var(--ov-gold);
-   font-size: 24px;
- }
- 
- .sidebar-menu {
-   list-style: none;
-   padding: 0;
-   margin: 0;
-   flex-grow: 1;
- }
- 
- .sidebar-menu li {
-   margin-bottom: 5px;
- }
- 
- .sidebar-menu a {
-   display: flex;
-   align-items: center;
-   padding: 12px 30px;
-   color: var(--ov-text-muted);
-   text-decoration: none;
-   font-size: 15px;
-   font-family: var(--font-typewriter);
-   border-left: 3px solid transparent;
-   transition: all 0.2s;
- }
- 
- .sidebar-menu li.active a,
- .sidebar-menu a:hover {
-   color: var(--ov-gold);
-   background: linear-gradient(90deg, rgba(212, 175, 55, 0.1) 0%, transparent 100%);
-   border-left-color: var(--ov-gold);
- }
- 
- .sidebar-footer {
-   padding: 0 20px;
- }
- 
- .footer-stamp {
-   border: 1px solid var(--ov-border-faint);
-   padding: 15px;
-   text-align: center;
-   border-radius: 4px;
-   margin-bottom: 20px;
- }
- 
- .footer-stamp small {
-   display: block;
-   font-size: 9px;
-   color: var(--ov-text-muted);
-   text-transform: uppercase;
-   letter-spacing: 1px;
- }
- 
- .footer-stamp span {
-   display: block;
-   font-family: var(--font-typewriter);
-   color: var(--ov-gold);
-   font-size: 13px;
-   margin: 5px 0;
- }
- 
- .lang-switch {
-   display: flex;
-   border: 1px solid var(--ov-border-light);
-   border-radius: 4px;
-   overflow: hidden;
- }
- 
- .lang-switch button {
-   flex: 1;
-   background: transparent;
-   border: none;
-   color: var(--ov-text-muted);
-   padding: 8px 0;
-   font-size: 12px;
-   cursor: pointer;
- }
- 
- .lang-switch button.active {
-   color: var(--ov-gold);
-   background: rgba(212, 175, 55, 0.05);
- }
- 
- /* ====================
-    Main Content Area
-    ==================== */
- #main-content {
-   margin-left: 240px;
-   padding: 40px 60px;
-   max-width: 1200px;
- }
- 
- #objectverse-hero {
-   margin-bottom: 40px;
-   position: relative;
- }
- 
- #objectverse-hero h1 {
-   font-family: var(--font-typewriter);
-   font-size: 42px;
-   color: var(--ov-text-main);
-   margin: 0 0 10px 0;
-   letter-spacing: -0.5px;
- }
- 
- .hero-kicker {
-   font-size: 18px;
-   color: var(--ov-gold);
-   font-style: italic;
-   font-family: var(--font-serif);
-   margin: 0;
- }
- 
- .hero-kicker span {
-   font-size: 14px;
-   font-style: normal;
-   color: #A89B84 !important;
-   font-family: var(--font-sans);
- }
- 
- .hero-badges {
-   display: flex;
-   gap: 15px;
-   margin-top: 25px;
- }
- 
- .hero-badges span {
-   border: 1px solid var(--ov-border-light);
-   padding: 6px 16px;
-   border-radius: 20px;
-   font-size: 13px;
-   color: var(--ov-text-muted);
-   font-family: var(--font-typewriter);
-   display: flex;
-   align-items: center;
-   gap: 6px;
- }
- 
- .content-section {
-   margin-bottom: 30px;
-   gap: 30px !important;
- }
- 
- .archive-panel {
-   background: var(--ov-bg-panel) !important;
-   border: 1px solid var(--ov-border-faint) !important;
-   border-radius: 8px;
-   padding: 25px;
-   position: relative;
- }
- 
- /* Gradio Overrides */
- .gradio-container .block,
- .gradio-container .form,
- .gradio-container .box {
-   background: transparent !important;
-   border: none !important;
-   box-shadow: none !important;
- }
- 
- .gradio-container label, .gradio-container span.svelte-1gfknul {
-   color: var(--ov-text-muted) !important;
-   font-family: var(--font-typewriter);
- }
- 
- .gradio-container input, .gradio-container textarea {
-   background: var(--ov-bg-input) !important;
-   border: 1px solid var(--ov-border-light) !important;
-   border-radius: 4px !important;
-   color: var(--ov-text-main) !important;
-   font-family: var(--font-sans) !important;
- }
- 
- .gradio-container input:focus, .gradio-container textarea:focus {
-   border-color: var(--ov-gold) !important;
-   box-shadow: none !important;
- }
- 
- /* Upload Box */
- #object-upload {
-   border: 2px dashed var(--ov-border-light) !important;
-   background: transparent !important;
-   border-radius: 8px;
-   padding: 40px 20px;
-   text-align: center;
-   min-height: 180px;
-   display: flex;
-   align-items: center;
-   justify-content: center;
- }
- 
- .or-divider {
-   text-align: center;
-   position: relative;
-   margin: 20px 0;
- }
- 
- .or-divider::before {
-   content: "";
-   position: absolute;
-   left: 0; right: 0; top: 50%;
-   height: 1px;
-   background: var(--ov-border-faint);
-   z-index: 1;
- }
- 
- .or-divider span {
-   background: var(--ov-bg-panel);
-   padding: 0 15px;
-   position: relative;
-   z-index: 2;
-   color: var(--ov-text-muted);
-   font-family: var(--font-typewriter);
-   font-size: 14px;
- }
- 
- /* Personality Mode Radio */
- .mode-header {
-   font-family: var(--font-typewriter);
-   color: var(--ov-text-main);
-   margin-bottom: 15px;
-   display: flex;
-   align-items: center;
-   gap: 10px;
- }
- 
- .mode-header small {
-   color: var(--ov-text-muted);
-   font-family: var(--font-sans);
- }
- 
- #personality-mode .wrap {
-   display: flex !important;
-   gap: 10px !important;
-   flex-wrap: wrap !important;
- }
- 
- #personality-mode label {
-   flex: 1;
-   background: transparent !important;
-   border: 1px solid var(--ov-border-light) !important;
-   border-radius: 6px !important;
-   padding: 15px 10px !important;
-   text-align: center;
-   cursor: pointer;
-   transition: all 0.2s;
- }
- 
- #personality-mode label span {
-   display: block;
-   font-family: var(--font-typewriter);
-   color: var(--ov-text-main) !important;
-   font-size: 14px;
- }
- 
- #personality-mode label:has(input:checked) {
-   border-color: var(--ov-gold) !important;
-   background: rgba(212, 175, 55, 0.05) !important;
-   box-shadow: 0 0 0 1px var(--ov-gold) inset;
- }
- 
- #personality-mode label:has(input:checked) span {
-   color: var(--ov-gold-bright) !important;
- }
- 
- /* Wake Button */
- #wake-button {
-   background: linear-gradient(180deg, #d8ac54 0%, #a67c2d 100%) !important;
-   border: none !important;
-   border-radius: 4px !important;
-   color: var(--ov-text-dark) !important;
-   font-family: var(--font-typewriter);
-   font-size: 20px !important;
-   font-weight: bold;
-   padding: 20px !important;
-   margin-top: 25px;
-   box-shadow: inset 0 1px 1px rgba(255,255,255,0.3), 0 4px 15px rgba(0,0,0,0.5) !important;
-   text-shadow: 0 1px 0 rgba(255,255,255,0.2);
-   transition: all 0.2s;
- }
- 
- #wake-button:hover {
-   filter: brightness(1.1);
-   transform: translateY(-1px);
- }
- 
- /* How it works */
- .how-it-works {
-   display: flex;
-   gap: 20px;
-   margin-top: 40px;
-   padding-top: 30px;
-   border-top: 1px dashed var(--ov-border-faint);
- }
- 
- .step {
-   flex: 1;
-   position: relative;
- }
- 
- .step-num {
-   position: absolute;
-   top: -10px; left: -10px;
-   background: var(--ov-bg);
-   border: 1px solid var(--ov-border-light);
-   color: var(--ov-gold);
-   font-family: var(--font-typewriter);
-   font-size: 12px;
-   padding: 2px 8px;
- }
- 
- .step-text strong {
-   display: block;
-   color: var(--ov-text-main);
-   font-family: var(--font-typewriter);
-   font-size: 14px;
-   margin-top: 15px;
- }
- 
- .step-text small {
-   display: block;
-   color: var(--ov-text-muted);
-   font-size: 12px;
-   margin-bottom: 8px;
- }
- 
- .step-text p {
-   color: var(--ov-text-muted);
-   font-size: 13px;
-   line-height: 1.4;
-   margin: 0;
- }
- 
- /* Example Objects Panel */
- .example-header {
-   display: flex;
-   align-items: center;
-   gap: 15px;
-   margin-bottom: 20px;
-   border-bottom: 1px solid var(--ov-border-faint);
-   padding-bottom: 15px;
- }
- 
- .example-header strong {
-   display: block;
-   font-family: var(--font-typewriter);
-   font-size: 16px;
-   font-weight: normal;
- }
- 
- .example-header span {
-   color: var(--ov-text-muted);
-   font-size: 13px;
- }
- 
- button.example-card {
-   background: rgba(22, 21, 19, 0.8) !important;
-   border: 1px solid var(--ov-border-faint) !important;
-   border-radius: 4px !important;
-   color: var(--ov-text-main) !important;
-   text-align: left !important;
-   padding: 15px !important;
-   margin-bottom: 12px !important;
-   font-family: var(--font-typewriter) !important;
-   display: block;
-   width: 100%;
-   transition: border-color 0.2s;
- }
- 
- button.example-card:hover {
-   border-color: var(--ov-gold) !important;
- }
- 
- .view-more {
-   display: block;
-   text-align: right;
-   color: var(--ov-gold);
-   text-decoration: none;
-   font-family: var(--font-typewriter);
-   font-size: 14px;
-   margin-top: 15px;
- }
- 
- /* Other Panels Formatting */
- .panel-header h2 {
-   font-family: var(--font-typewriter);
-   font-size: 24px;
-   color: var(--ov-text-main);
-   margin: 0 0 5px 0;
- }
- 
- .panel-header {
-   border-bottom: 1px solid var(--ov-border-faint);
-   padding-bottom: 15px;
-   margin-bottom: 20px;
- }
- 
- .panel-header > span {
-   background: transparent;
-   border: none;
-   color: var(--ov-gold) !important;
-   font-family: var(--font-typewriter);
-   font-size: 18px;
-   padding: 0;
- }
- 
- /* Markdown & Typography */
- #diary-output {
-   font-family: var(--font-serif) !important;
-   font-size: 18px;
-   line-height: 1.8;
-   color: #D6D1C4 !important;
- }
- 
- #diary-output h3 {
-   font-family: var(--font-typewriter);
-   color: var(--ov-gold);
-   text-transform: uppercase;
-   font-size: 16px;
- }
- 
- .archive-empty {
-   text-align: center;
-   padding: 40px;
-   border: 1px dashed var(--ov-border-light);
- }
- 
- .archive-empty h3 {
-   font-family: var(--font-typewriter);
- }
- 
- /* Responsive */
- @media (max-width: 980px) {
-   #app-container {
-     flex-direction: column;
-   }
-   #sidebar {
-     position: static;
-     width: 100% !important;
-     max-width: 100% !important;
-     height: auto;
-     padding: 20px;
-     border-right: none;
-     border-bottom: 1px solid var(--ov-border-faint);
-   }
-   #main-content {
-     margin-left: 0;
-     padding: 20px;
-   }
-   .content-section {
-     flex-direction: column !important;
-   }
-   .split-section {
-     flex-direction: column !important;
-   }
- }
-
- @media (max-width: 600px) {
-   #main-content {
-     padding: 15px !important;
-   }
-   #objectverse-hero h1 {
-     font-size: 28px !important;
-     word-break: break-word;
-   }
-   .hero-kicker {
-     font-size: 15px !important;
-   }
-   #personality-mode label {
-     flex: 1 1 45% !important;
-     padding: 10px 5px !important;
-   }
-   .sidebar-menu {
-     display: flex;
-     flex-wrap: wrap;
-     gap: 5px;
-   }
-   .sidebar-menu li {
-     margin-bottom: 0;
-   }
-   .sidebar-menu a {
-     padding: 8px 10px;
-     font-size: 13px;
-     border-left: none;
-     border-bottom: 2px solid transparent;
-   }
-   .sidebar-menu li.active a {
-     border-bottom-color: var(--ov-gold);
-     border-left: none;
-     background: rgba(212, 175, 55, 0.1);
-   }
-   .lang-switch {
-     margin-top: 10px;
-   }
-   .how-it-works {
-     flex-direction: column;
-     gap: 20px;
-   }
-   .hero-badges {
-     flex-wrap: wrap;
-   }
-   .hero-badges span {
-     flex: 1 1 100%;
-     justify-content: center;
-   }
- }
+@import url('https://fonts.googleapis.com/css2?family=Space+Mono:ital,wght@0,400;0,700;1,400&family=Courier+Prime:ital,wght@0,400;0,700;1,400&display=swap');
+
+:root {
+  --ov-bg: #161513;
+  --ov-bg-panel: rgba(30, 28, 25, 0.72);
+  --ov-bg-input: #1b1a18;
+  --ov-border-faint: rgba(212, 175, 55, 0.15);
+  --ov-border-light: rgba(212, 175, 55, 0.34);
+  --ov-text-main: #e6e1d3;
+  --ov-text-soft: #d6d1c4;
+  --ov-text-muted: #8b8678;
+  --ov-text-dark: #2a261f;
+  --ov-gold: #d4af37;
+  --ov-gold-bright: #f5d061;
+  --font-typewriter: 'Courier Prime', 'Space Mono', 'Courier New', monospace;
+  --font-sans: 'Inter', -apple-system, BlinkMacSystemFont, 'Segoe UI', sans-serif;
+  --font-serif: Georgia, serif;
+}
+
+.lang-zh,
+.auto-zh {
+  display: none !important;
+}
+
+html[data-ov-lang="zh"] .lang-zh,
+html[data-ov-lang="zh"] .auto-zh {
+  display: inline !important;
+}
+
+html[data-ov-lang="zh"] .lang-zh.block {
+  display: block !important;
+}
+
+html,
+body,
+gradio-app {
+  background: var(--ov-bg);
+  color: var(--ov-text-main);
+  min-height: 100%;
+  margin: 0;
+  overflow-x: hidden;
+}
+
+body::before {
+  content: "";
+  position: fixed;
+  inset: 0;
+  background-image: url("data:image/svg+xml,%3Csvg viewBox='0 0 200 200' xmlns='http://www.w3.org/2000/svg'%3E%3Cfilter id='noiseFilter'%3E%3CfeTurbulence type='fractalNoise' baseFrequency='0.85' numOctaves='3' stitchTiles='stitch'/%3E%3C/filter%3E%3Crect width='100%25' height='100%25' filter='url(%23noiseFilter)' opacity='0.03'/%3E%3C/svg%3E");
+  pointer-events: none;
+  z-index: 9999;
+}
+
+.gradio-container {
+  width: 100% !important;
+  max-width: 100% !important;
+  min-height: 100vh !important;
+  padding: 0 !important;
+  background: transparent !important;
+  color: var(--ov-text-main) !important;
+  font-family: var(--font-sans);
+}
+
+footer,
+.footer {
+  display: none !important;
+}
+
+#objectverse-app,
+#app-container,
+#main-content,
+.archive-panel {
+  color: var(--ov-text-main);
+}
+
+#app-container {
+  max-width: 1180px;
+  margin: 0 auto !important;
+  padding: 36px 24px 56px;
+  gap: 24px !important;
+}
+
+#objectverse-hero {
+  display: flex;
+  justify-content: space-between;
+  align-items: flex-start;
+  gap: 24px;
+  padding-bottom: 22px;
+  border-bottom: 1px solid var(--ov-border-faint);
+}
+
+.hero-copy h1 {
+  margin: 6px 0 8px;
+  color: var(--ov-text-main) !important;
+  font-family: var(--font-typewriter);
+  font-size: clamp(32px, 5vw, 48px);
+  line-height: 1.05;
+  letter-spacing: 0;
+}
+
+.hero-kicker {
+  margin: 0;
+  color: var(--ov-gold) !important;
+  font-family: var(--font-serif);
+  font-size: 18px;
+  font-style: italic;
+}
+
+.hero-feature {
+  max-width: 680px;
+  margin: 12px 0 0;
+  color: var(--ov-text-soft) !important;
+  font-size: 15px;
+  line-height: 1.65;
+}
+
+.archive-label {
+  color: var(--ov-gold) !important;
+  font-family: var(--font-typewriter);
+  font-size: 12px;
+  letter-spacing: 0;
+  text-transform: uppercase;
+}
+
+.top-controls {
+  display: flex;
+  align-items: center;
+  gap: 12px;
+  min-width: 176px;
+  color: var(--ov-text-muted);
+  font-family: var(--font-typewriter);
+  font-size: 12px;
+}
+
+.segmented-control {
+  display: grid;
+  grid-template-columns: 1fr 1fr;
+  min-width: 92px;
+  border: 1px solid var(--ov-border-light);
+  border-radius: 4px;
+  overflow: hidden;
+}
+
+.segmented-control button {
+  min-height: 34px;
+  padding: 0 12px;
+  background: transparent;
+  border: 0;
+  border-right: 1px solid var(--ov-border-faint);
+  color: var(--ov-text-muted);
+  font-family: var(--font-typewriter);
+  font-size: 12px;
+  cursor: pointer;
+}
+
+.segmented-control button:last-child {
+  border-right: 0;
+}
+
+.segmented-control button:hover,
+.segmented-control button:focus,
+.segmented-control button.active {
+  color: var(--ov-gold);
+  background: rgba(212, 175, 55, 0.08);
+  outline: none;
+}
+
+.content-section {
+  display: flex !important;
+  gap: 24px !important;
+  margin: 0 0 2px;
+}
+
+.archive-panel {
+  position: relative;
+  padding: 22px;
+  background: var(--ov-bg-panel) !important;
+  border: 1px solid var(--ov-border-faint) !important;
+  border-radius: 8px;
+}
+
+.gradio-container .block,
+.gradio-container .form,
+.gradio-container .box {
+  background: transparent !important;
+  border: none !important;
+  box-shadow: none !important;
+}
+
+.gradio-container label,
+.gradio-container span.svelte-1gfknul {
+  color: var(--ov-text-muted) !important;
+  font-family: var(--font-typewriter);
+}
+
+.gradio-container input,
+.gradio-container textarea {
+  background: var(--ov-bg-input) !important;
+  border: 1px solid var(--ov-border-light) !important;
+  border-radius: 4px !important;
+  color: var(--ov-text-main) !important;
+  font-family: var(--font-sans) !important;
+}
+
+.gradio-container input:focus,
+.gradio-container textarea:focus {
+  border-color: var(--ov-gold) !important;
+  box-shadow: none !important;
+}
+
+#object-upload {
+  display: flex;
+  align-items: center;
+  justify-content: center;
+  min-height: 176px;
+  padding: 34px 20px;
+  border: 2px dashed var(--ov-border-light) !important;
+  border-radius: 8px;
+  background: transparent !important;
+  color: var(--ov-text-muted) !important;
+  text-align: center;
+}
+
+#object-upload :is(label, span, p, div, button) {
+  color: var(--ov-text-muted) !important;
+}
+
+.or-divider {
+  position: relative;
+  margin: 18px 0;
+  text-align: center;
+}
+
+.or-divider::before {
+  content: "";
+  position: absolute;
+  top: 50%;
+  right: 0;
+  left: 0;
+  height: 1px;
+  background: var(--ov-border-faint);
+}
+
+.or-divider span {
+  position: relative;
+  z-index: 1;
+  padding: 0 14px;
+  background: var(--ov-bg-panel);
+  color: var(--ov-text-muted);
+  font-family: var(--font-typewriter);
+  font-size: 13px;
+}
+
+.mode-header {
+  display: flex;
+  align-items: center;
+  gap: 8px;
+  margin: 18px 0 12px;
+  color: var(--ov-text-main) !important;
+  font-family: var(--font-typewriter);
+}
+
+.mode-header .lang-zh {
+  color: var(--ov-text-muted);
+  font-family: var(--font-sans);
+  font-size: 12px;
+}
+
+#personality-mode .wrap {
+  display: flex !important;
+  flex-wrap: wrap !important;
+  gap: 10px !important;
+}
+
+#personality-mode label {
+  flex: 1 1 120px;
+  min-height: 48px;
+  padding: 13px 10px !important;
+  border: 1px solid var(--ov-border-light) !important;
+  border-radius: 6px !important;
+  background: transparent !important;
+  text-align: center;
+  cursor: pointer;
+}
+
+#personality-mode label span {
+  display: block;
+  color: var(--ov-text-main) !important;
+  font-family: var(--font-typewriter);
+  font-size: 14px;
+}
+
+#personality-mode label:has(input:checked) {
+  border-color: var(--ov-gold) !important;
+  background: rgba(212, 175, 55, 0.06) !important;
+  box-shadow: 0 0 0 1px var(--ov-gold) inset;
+}
+
+#personality-mode label:has(input:checked) span {
+  color: var(--ov-gold-bright) !important;
+}
+
+#wake-button {
+  width: 100%;
+  margin-top: 22px;
+  padding: 18px !important;
+  border: 0 !important;
+  border-radius: 4px !important;
+  background: linear-gradient(180deg, #d8ac54 0%, #a67c2d 100%) !important;
+  color: var(--ov-text-dark) !important;
+  font-family: var(--font-typewriter);
+  font-size: 18px !important;
+  font-weight: 700;
+  box-shadow: inset 0 1px 1px rgba(255, 255, 255, 0.28), 0 4px 14px rgba(0, 0, 0, 0.34) !important;
+}
+
+#wake-button:hover {
+  filter: brightness(1.08);
+  transform: translateY(-1px);
+}
+
+.example-header {
+  margin-bottom: 18px;
+  padding-bottom: 14px;
+  border-bottom: 1px solid var(--ov-border-faint);
+}
+
+.example-header strong {
+  display: block;
+  color: var(--ov-text-main) !important;
+  font-family: var(--font-typewriter);
+  font-size: 16px;
+  font-weight: 400;
+}
+
+.example-header span {
+  color: var(--ov-text-muted) !important;
+  font-size: 13px;
+}
+
+button.example-card {
+  display: block;
+  width: 100%;
+  margin-bottom: 10px !important;
+  padding: 14px !important;
+  border: 1px solid var(--ov-border-faint) !important;
+  border-radius: 4px !important;
+  background: var(--ov-bg-input) !important;
+  color: var(--ov-text-main) !important;
+  font-family: var(--font-typewriter) !important;
+  text-align: left !important;
+  transition: border-color 0.2s, background 0.2s;
+}
+
+button.example-card:hover,
+button.example-card:focus {
+  border-color: var(--ov-gold) !important;
+}
+
+button.example-card * {
+  color: inherit !important;
+}
+
+.panel-header {
+  display: grid;
+  grid-template-columns: auto 1fr;
+  gap: 14px;
+  margin-bottom: 18px;
+  padding-bottom: 14px;
+  border-bottom: 1px solid var(--ov-border-faint);
+}
+
+.panel-header > span {
+  color: var(--ov-gold) !important;
+  font-family: var(--font-typewriter);
+  font-size: 16px;
+}
+
+.panel-header h2 {
+  margin: 0 0 4px;
+  color: var(--ov-text-main) !important;
+  font-family: var(--font-typewriter);
+  font-size: 22px;
+  line-height: 1.2;
+}
+
+.panel-header h2 small,
+.panel-header p {
+  color: var(--ov-text-muted) !important;
+}
+
+.panel-header p {
+  margin: 0;
+  font-size: 13px;
+  line-height: 1.45;
+}
+
+#diary-output,
+.diary-entry {
+  color: var(--ov-text-soft) !important;
+  font-family: var(--font-serif) !important;
+  font-size: 17px;
+  line-height: 1.75;
+}
+
+.diary-entry h2,
+#diary-output h2,
+#diary-output h3 {
+  margin: 0 0 14px;
+  color: var(--ov-gold) !important;
+  font-family: var(--font-typewriter);
+  font-size: 17px;
+  line-height: 1.3;
+  text-transform: uppercase;
+}
+
+.diary-entry p {
+  margin: 0;
+}
+
+.zh-helper {
+  margin-top: 18px;
+  padding-top: 14px;
+  border-top: 1px dashed var(--ov-border-faint);
+  color: var(--ov-text-muted) !important;
+  font-family: var(--font-sans);
+  font-size: 14px;
+  line-height: 1.7;
+}
+
+.archive-empty,
+.objectverse-placeholder {
+  padding: 34px 24px;
+  border: 1px dashed var(--ov-border-light);
+  color: var(--ov-text-muted) !important;
+  text-align: center;
+}
+
+.archive-empty h3,
+.objectverse-placeholder strong {
+  display: block;
+  margin: 8px 0;
+  color: var(--ov-text-main) !important;
+  font-family: var(--font-typewriter);
+}
+
+.archive-empty p,
+.objectverse-placeholder p {
+  margin: 6px 0 0;
+  color: var(--ov-text-muted) !important;
+}
+
+.object-file-card,
+.trace-card,
+.archive-error,
+.objectverse-card {
+  color: var(--ov-text-main) !important;
+}
+
+.object-file-card h3,
+.trace-card strong,
+.archive-error strong,
+.objectverse-card h2 {
+  color: var(--ov-text-main) !important;
+}
+
+.file-meta {
+  display: flex;
+  flex-wrap: wrap;
+  gap: 8px;
+  margin-bottom: 12px;
+}
+
+.file-meta span,
+.object-name,
+.object-file-card dt,
+.object-file-card p,
+.object-file-card li,
+.trace-card p,
+.archive-error p,
+.card-kicker,
+.card-object,
+.card-cn {
+  color: var(--ov-text-muted) !important;
+}
+
+.object-file-card dl {
+  display: grid;
+  gap: 12px;
+  margin: 18px 0;
+}
+
+.object-file-card dl > div {
+  padding-bottom: 10px;
+  border-bottom: 1px solid var(--ov-border-faint);
+}
+
+.object-file-card dt {
+  margin-bottom: 4px;
+  font-family: var(--font-typewriter);
+  font-size: 12px;
+  text-transform: uppercase;
+}
+
+.object-file-card dd {
+  margin: 0;
+  color: var(--ov-text-soft) !important;
+}
+
+.feature-list strong,
+.card-quote {
+  color: var(--ov-text-soft) !important;
+}
+
+.feature-list ul {
+  margin: 8px 0 0;
+  padding-left: 20px;
+}
+
+.complaint,
+.card-stamp {
+  color: var(--ov-gold) !important;
+}
+
+.file-tags,
+.card-tags {
+  display: flex;
+  flex-wrap: wrap;
+  gap: 8px;
+  margin-top: 16px;
+}
+
+.file-tags span,
+.card-tags span {
+  padding: 4px 8px;
+  border: 1px solid var(--ov-border-light) !important;
+  border-radius: 999px;
+  color: var(--ov-gold-bright) !important;
+  font-family: var(--font-typewriter);
+  font-size: 12px;
+}
+
+.objectverse-card {
+  max-width: 520px;
+  padding: 22px;
+  border: 1px solid var(--ov-border-light);
+  border-radius: 8px;
+  background: rgba(22, 21, 19, 0.7);
+}
+
+.card-header {
+  display: flex;
+  justify-content: space-between;
+  gap: 16px;
+  margin-bottom: 16px;
+}
+
+.card-header h2 {
+  margin: 4px 0 0;
+  font-family: var(--font-typewriter);
+}
+
+.card-quote {
+  font-family: var(--font-serif);
+  font-size: 17px;
+  line-height: 1.65;
+}
+
+.card-cn {
+  margin-top: 12px;
+  padding-top: 12px;
+  border-top: 1px dashed var(--ov-border-faint);
+}
+
+.quiet-button {
+  border: 1px solid var(--ov-border-light) !important;
+  border-radius: 4px !important;
+  background: transparent !important;
+  color: var(--ov-gold) !important;
+  font-family: var(--font-typewriter) !important;
+}
+
+.developer-details {
+  border: 1px solid var(--ov-border-faint) !important;
+  border-radius: 8px !important;
+  background: rgba(30, 28, 25, 0.42) !important;
+}
+
+.developer-json-grid {
+  gap: 16px !important;
+}
+
+.gradio-container :is(summary, pre, code) {
+  color: var(--ov-text-main) !important;
+}
+
+.gradio-container :is(.json-holder, .json-holder *, .chatbot, .chatbot *) {
+  color: var(--ov-text-main) !important;
+}
+
+@media (max-width: 980px) {
+  #app-container {
+    padding: 28px 18px 44px;
+  }
+
+  #objectverse-hero,
+  .content-section,
+  .split-section,
+  .developer-json-grid {
+    flex-direction: column !important;
+  }
+
+  .top-controls {
+    width: 100%;
+    justify-content: space-between;
+  }
+}
+
+@media (max-width: 600px) {
+  #app-container {
+    padding: 22px 14px 36px;
+  }
+
+  .archive-panel {
+    padding: 18px;
+  }
+
+  .hero-copy h1 {
+    font-size: 30px;
+    word-break: break-word;
+  }
+
+  .hero-kicker {
+    font-size: 15px;
+  }
+
+  .hero-feature {
+    font-size: 14px;
+    line-height: 1.55;
+  }
+
+  #personality-mode label {
+    flex: 1 1 45% !important;
+    padding: 10px 6px !important;
+  }
+}

src/utils/json_repair.py

@@ -1,4 +1,4 @@
-"""JSON repair placeholder for later model integration."""
+"""Small JSON object extraction helpers for model output."""
 
 from __future__ import annotations
 
@@ -14,17 +14,61 @@ def parse_json_object(raw: str) -> dict[str, Any]:
 
 
 def _extract_json_object(raw: str) -> str:
-    clean = raw.strip()
-    if clean.startswith("```"):
-        clean = clean.strip("`").strip()
-        if clean.lower().startswith("json"):
-            clean = clean[4:].strip()
+    clean = _strip_markdown_fence(raw.strip())
 
-    if clean.startswith("{") and clean.endswith("}"):
+    candidate = _scan_json_object(clean)
+    if candidate is None:
+        raise ValueError("No JSON object found.")
+    return candidate
+
+
+def _strip_markdown_fence(clean: str) -> str:
+    if not clean.startswith("```"):
         return clean
 
+    lines = clean.splitlines()
+    if not lines:
+        return clean
+
+    if lines[0].strip().startswith("```"):
+        lines = lines[1:]
+    if lines and lines[-1].strip().startswith("```"):
+        lines = lines[:-1]
+    return "\n".join(lines).strip()
+
+
+def _scan_json_object(clean: str) -> str | None:
     start = clean.find("{")
-    end = clean.rfind("}")
-    if start == -1 or end == -1 or end <= start:
-        raise ValueError("No JSON object found.")
-    return clean[start : end + 1]
+    if start == -1:
+        return None
+
+    stack: list[str] = []
+    in_string = False
+    escaped = False
+
+    for index, char in enumerate(clean[start:], start=start):
+        if in_string:
+            if escaped:
+                escaped = False
+            elif char == "\\":
+                escaped = True
+            elif char == '"':
+                in_string = False
+            continue
+
+        if char == '"':
+            in_string = True
+        elif char == "{":
+            stack.append("}")
+        elif char == "[":
+            stack.append("]")
+        elif char in {"}", "]"}:
+            if not stack or stack[-1] != char:
+                return None
+            stack.pop()
+            if not stack:
+                return clean[start : index + 1]
+
+    if in_string or not stack:
+        return None
+    return clean[start:] + "".join(reversed(stack))

tests/test_dataset_tooling.py

@@ -5,12 +5,13 @@ from __future__ import annotations
 import json
 import tempfile
 import unittest
+from collections import Counter
 from pathlib import Path
 
 from scripts.export_traces import export_trace_jsonl
 from scripts.generate_dataset import build_sft_records, write_sft_jsonl
 from scripts.generate_sample_traces import generate_sample_traces
-from scripts.prepare_curated_dataset import build_curated_records, write_jsonl
+from scripts.prepare_curated_dataset import MODES, build_curated_records, write_jsonl
 from src.models.schema import TraceRecord
 
 
@@ -48,6 +49,25 @@ class DatasetToolingTest(unittest.TestCase):
         self.assertIn("persona", assistant_payload)
         self.assertIn("diary", assistant_payload)
 
+    def test_build_curated_v2_records_has_broader_balanced_coverage(self) -> None:
+        records = build_curated_records(200, version="v2")
+        object_names = [
+            record["object_understanding"]["object"]["name"]
+            for record in records
+        ]
+        mode_counts = Counter(record["mode"] for record in records)
+        object_mode_pairs = {(name, record["mode"]) for name, record in zip(object_names, records)}
+        assistant_payload = json.loads(records[0]["messages"][2]["content"])
+
+        self.assertEqual(len(records), 200)
+        self.assertEqual(records[0]["source"], "objectverse-diary-synthetic-curated-v2")
+        self.assertGreaterEqual(len(set(object_names)), 40)
+        self.assertEqual(mode_counts, Counter({mode: 40 for mode in MODES}))
+        self.assertEqual(len(object_mode_pairs), 200)
+        self.assertIn("scene_detail", records[0])
+        self.assertIn("persona", assistant_payload)
+        self.assertIn("diary", assistant_payload)
+
     def test_write_curated_jsonl(self) -> None:
         with tempfile.TemporaryDirectory() as tmp_dir:
             output_path = Path(tmp_dir) / "curated.jsonl"

tests/test_finetune_lora_tooling.py

@@ -10,6 +10,41 @@ from pathlib import Path
 from scripts import finetune_lora
 
 
+class FakeTokenizer:
+    pad_token = "<pad>"
+    eos_token = "</s>"
+
+    def apply_chat_template(
+        self,
+        messages: list[dict[str, str]],
+        *,
+        tokenize: bool,
+        add_generation_prompt: bool,
+    ) -> str:
+        text = "".join(
+            f"<{message['role']}>{message['content']}</{message['role']}>"
+            for message in messages
+        )
+        if add_generation_prompt:
+            text += "<assistant>"
+        return text
+
+    def __call__(
+        self,
+        text: str,
+        *,
+        truncation: bool,
+        max_length: int,
+        padding: bool,
+        add_special_tokens: bool = False,
+    ) -> dict[str, list[int]]:
+        del padding, add_special_tokens
+        ids = [ord(character) % 251 + 1 for character in text]
+        if truncation:
+            ids = ids[:max_length]
+        return {"input_ids": ids, "attention_mask": [1] * len(ids)}
+
+
 def _valid_record() -> dict[str, object]:
     return {
         "id": "sft-preview-0001",
@@ -55,9 +90,38 @@ class FinetuneLoraToolingTest(unittest.TestCase):
         self.assertEqual(config.lora_alpha, 32)
         self.assertEqual(config.lora_dropout, 0.05)
         self.assertEqual(config.max_steps, 80)
+        self.assertEqual(config.num_train_epochs, 3.0)
+        self.assertEqual(config.per_device_train_batch_size, 1)
+        self.assertEqual(config.gradient_accumulation_steps, 4)
+        self.assertEqual(config.eval_ratio, 0.1)
+        self.assertTrue(config.assistant_only_loss)
         self.assertIn("q_proj", config.target_modules)
         self.assertIn("down_proj", config.target_modules)
 
+    def test_training_config_serializes_v2_experiment_settings(self) -> None:
+        config = finetune_lora.TrainingConfig(
+            max_steps=0,
+            num_train_epochs=4.0,
+            per_device_train_batch_size=2,
+            gradient_accumulation_steps=8,
+            eval_ratio=0.2,
+            eval_steps=25,
+            lora_r=32,
+            lora_alpha=64,
+            assistant_only_loss=False,
+        )
+
+        payload = config.as_remote_dict()
+
+        self.assertEqual(payload["num_train_epochs"], 4.0)
+        self.assertEqual(payload["per_device_train_batch_size"], 2)
+        self.assertEqual(payload["gradient_accumulation_steps"], 8)
+        self.assertEqual(payload["eval_ratio"], 0.2)
+        self.assertEqual(payload["eval_steps"], 25)
+        self.assertEqual(payload["lora_r"], 32)
+        self.assertEqual(payload["lora_alpha"], 64)
+        self.assertFalse(payload["assistant_only_loss"])
+
     def test_dry_run_does_not_call_remote_runner(self) -> None:
         with tempfile.TemporaryDirectory() as tmp_dir:
             path = Path(tmp_dir) / "records.jsonl"
@@ -80,6 +144,44 @@ class FinetuneLoraToolingTest(unittest.TestCase):
         self.assertEqual(summary["mode"], "dry-run")
         self.assertEqual(summary["record_count"], 1)
         self.assertEqual(summary["base_model"], "Qwen/Qwen2.5-1.5B-Instruct")
+        self.assertEqual(summary["train_record_count"], 1)
+        self.assertEqual(summary["eval_record_count"], 0)
+
+    def test_dry_run_reports_eval_split_for_larger_datasets(self) -> None:
+        records = [_valid_record() for _ in range(20)]
+        summary = finetune_lora._dry_run_summary(
+            Path("records.jsonl"),
+            records,
+            finetune_lora.TrainingConfig(eval_ratio=0.2),
+        )
+
+        self.assertEqual(summary["train_record_count"], 16)
+        self.assertEqual(summary["eval_record_count"], 4)
+
+    def test_assistant_only_tokenization_masks_prompt_labels(self) -> None:
+        tokenized = finetune_lora._tokenize_training_example(
+            _valid_record(),
+            FakeTokenizer(),
+            max_length=512,
+            assistant_only_loss=True,
+        )
+
+        labels = tokenized["labels"]
+
+        self.assertIn(-100, labels)
+        self.assertTrue(any(label != -100 for label in labels))
+        first_unmasked = next(index for index, label in enumerate(labels) if label != -100)
+        self.assertGreater(first_unmasked, 0)
+
+    def test_full_loss_tokenization_keeps_all_labels(self) -> None:
+        tokenized = finetune_lora._tokenize_training_example(
+            _valid_record(),
+            FakeTokenizer(),
+            max_length=512,
+            assistant_only_loss=False,
+        )
+
+        self.assertNotIn(-100, tokenized["labels"])
 
 
 if __name__ == "__main__":

tests/test_json_repair.py

@@ -0,0 +1,34 @@
+"""Tests for tolerant JSON object extraction from model output."""
+
+from __future__ import annotations
+
+import unittest
+
+from src.utils.json_repair import parse_json_object
+
+
+class JsonRepairTest(unittest.TestCase):
+    def test_parses_complete_object_with_surrounding_text(self) -> None:
+        payload = parse_json_object('Here is the archive:\n{"name": "mug"}\nDone.')
+
+        self.assertEqual(payload, {"name": "mug"})
+
+    def test_repairs_missing_outer_closing_brace(self) -> None:
+        payload = parse_json_object(
+            """
+            {
+              "persona": {"object_name": "coffee mug"},
+              "diary": {"title": "Secret Diary - Day 310"}
+            """
+        )
+
+        self.assertEqual(payload["persona"]["object_name"], "coffee mug")
+        self.assertEqual(payload["diary"]["title"], "Secret Diary - Day 310")
+
+    def test_does_not_repair_unterminated_string(self) -> None:
+        with self.assertRaises(ValueError):
+            parse_json_object('{"name": "coffee mug}')
+
+
+if __name__ == "__main__":
+    unittest.main()

tests/test_llama_cpp_smoke.py

@@ -24,10 +24,7 @@ class LlamaCppSmokeTest(unittest.TestCase):
         fake_llama = FakeLlamaModel(
             [
                 """
-                {"persona":{"object_name":"coffee mug","character_name":"Mugworth","mood":"dry and suspicious","secret_fear":"being left empty forever","core_memory":"It remembers every late-night refill.","complaint":"I am treated like a ceramic fuel tank.","tags":["desk witness","warm archive","quiet judgment"]}}
-                """,
-                """
-                {"title":"Secret Diary - Day 418","english":"Today I held another bitter storm and called it service.","chinese":"今天我又装下一场苦涩风暴，并被称为有用。"}
+                {"persona":{"object_name":"coffee mug","character_name":"Mugworth","mood":"dry and suspicious","secret_fear":"being left empty forever","core_memory":"It remembers every late-night refill.","complaint":"I am treated like a ceramic fuel tank.","tags":["desk witness","warm archive","quiet judgment"]},"diary":{"title":"Secret Diary - Day 418","english":"Today I held another bitter storm and called it service.","chinese":"今天我又装下一场苦涩风暴，并被称为有用。"}}
                 """,
                 """
                 {"reply":"Mugworth: I saw another deadline dissolve into a coffee ring."}

tests/test_merge_lora_adapter.py

@@ -0,0 +1,46 @@
+"""Tests for LoRA merge helper tooling."""
+
+from __future__ import annotations
+
+import json
+import tempfile
+import unittest
+from pathlib import Path
+
+from scripts.merge_lora_adapter import plan_merge, validate_adapter_source
+
+
+class MergeLoraAdapterTest(unittest.TestCase):
+    def test_validate_local_adapter_requires_config_and_weights(self) -> None:
+        with tempfile.TemporaryDirectory() as tmp_dir:
+            adapter_dir = Path(tmp_dir)
+            (adapter_dir / "adapter_config.json").write_text("{}", encoding="utf-8")
+
+            with self.assertRaises(ValueError):
+                validate_adapter_source(adapter_dir, base_model="Qwen/Qwen2.5-1.5B-Instruct")
+
+    def test_plan_merge_dry_run_returns_summary_without_loading_model(self) -> None:
+        with tempfile.TemporaryDirectory() as tmp_dir:
+            adapter_dir = Path(tmp_dir) / "adapter"
+            adapter_dir.mkdir()
+            (adapter_dir / "adapter_config.json").write_text(
+                json.dumps({"base_model_name_or_path": "Qwen/Qwen2.5-1.5B-Instruct"}),
+                encoding="utf-8",
+            )
+            (adapter_dir / "adapter_model.safetensors").write_text("fake", encoding="utf-8")
+
+            summary = plan_merge(
+                base_model="Qwen/Qwen2.5-1.5B-Instruct",
+                adapter=adapter_dir,
+                output=Path(tmp_dir) / "merged",
+                dry_run=True,
+            )
+
+        self.assertTrue(summary["dry_run"])
+        self.assertFalse(summary["merged"])
+        self.assertEqual(summary["base_model"], "Qwen/Qwen2.5-1.5B-Instruct")
+        self.assertEqual(summary["adapter_type"], "local")
+
+
+if __name__ == "__main__":
+    unittest.main()

tests/test_publish_hf_dataset.py

@@ -0,0 +1,57 @@
+"""Tests for Hugging Face Dataset publishing helpers."""
+
+from __future__ import annotations
+
+import json
+import tempfile
+import unittest
+from pathlib import Path
+
+from scripts.prepare_curated_dataset import build_curated_records, write_jsonl
+from scripts.publish_hf_dataset import upload_dataset, validate_dataset_file
+
+
+class PublishHfDatasetTest(unittest.TestCase):
+    def test_validate_dataset_file_rejects_bad_assistant_json(self) -> None:
+        with tempfile.TemporaryDirectory() as tmp_dir:
+            dataset_file = Path(tmp_dir) / "bad.jsonl"
+            dataset_file.write_text(
+                json.dumps(
+                    {
+                        "id": "bad",
+                        "messages": [
+                            {"role": "system", "content": "system"},
+                            {"role": "user", "content": "user"},
+                            {"role": "assistant", "content": "not json"},
+                        ],
+                    }
+                )
+                + "\n",
+                encoding="utf-8",
+            )
+
+            with self.assertRaises(ValueError):
+                validate_dataset_file(dataset_file)
+
+    def test_upload_dataset_dry_run_returns_file_summary(self) -> None:
+        with tempfile.TemporaryDirectory() as tmp_dir:
+            dataset_file = Path(tmp_dir) / "curated_v2.jsonl"
+            write_jsonl(build_curated_records(5, version="v2"), dataset_file)
+
+            summary = upload_dataset(
+                dataset_file=dataset_file,
+                repo_id="qqyule/objectverse-diary-sft-curated",
+                path_in_repo="objectverse_sft_curated_v2.jsonl",
+                private=False,
+                commit_message="Dry run",
+                dry_run=True,
+            )
+
+        self.assertFalse(summary["uploaded"])
+        self.assertEqual(summary["repo_id"], "qqyule/objectverse-diary-sft-curated")
+        self.assertEqual(summary["path_in_repo"], "objectverse_sft_curated_v2.jsonl")
+        self.assertEqual(summary["record_count"], 5)
+
+
+if __name__ == "__main__":
+    unittest.main()

tests/test_publish_hf_gguf.py

@@ -0,0 +1,45 @@
+"""Tests for Hugging Face GGUF publishing helpers."""
+
+from __future__ import annotations
+
+import tempfile
+import unittest
+from pathlib import Path
+
+from scripts.publish_hf_gguf import upload_gguf, validate_gguf_file
+
+
+class PublishHfGgufTest(unittest.TestCase):
+    def test_validate_gguf_file_requires_gguf_suffix(self) -> None:
+        with tempfile.TemporaryDirectory() as tmp_dir:
+            model_file = Path(tmp_dir) / "model.bin"
+            model_file.write_text("fake", encoding="utf-8")
+
+            with self.assertRaises(ValueError):
+                validate_gguf_file(model_file)
+
+    def test_upload_gguf_dry_run_returns_file_summary(self) -> None:
+        with tempfile.TemporaryDirectory() as tmp_dir:
+            gguf_file = Path(tmp_dir) / "model.gguf"
+            gguf_file.write_bytes(b"fake-gguf")
+
+            summary = upload_gguf(
+                gguf_file=gguf_file,
+                repo_id="qqyule/objectverse-diary-qwen15b-lora",
+                path_in_repo="objectverse-diary-qwen15b-lora-v2-q4_k_m.gguf",
+                private=False,
+                commit_message="Dry run",
+                dry_run=True,
+            )
+
+        self.assertFalse(summary["uploaded"])
+        self.assertEqual(summary["repo_id"], "qqyule/objectverse-diary-qwen15b-lora")
+        self.assertEqual(
+            summary["path_in_repo"],
+            "objectverse-diary-qwen15b-lora-v2-q4_k_m.gguf",
+        )
+        self.assertEqual(summary["size_bytes"], 9)
+
+
+if __name__ == "__main__":
+    unittest.main()